This demo paper presents DataVerse, a web-based data collection application designed for research institutions. DataVerse treats multi-platform data collection as a reliable, scalable system rather than a single interface: collection tasks are modeled as independent, reusable actors, and user submissions are decoupled from execution as persisted job objects that can be queued, monitored, and stored independently. On top of this, DataVerse adds a governance layer built on hierarchical role-based access control (Super Admin, Admin, Editor, Viewer) and an approval-gated job lifecycle that enforces a separation of duties between the users who request a job and those who authorize its execution. This design keeps the interface responsive under long-running jobs, isolates failures to individual actors, and gives administrators an auditable record of how and why each job ran. We demonstrate DataVerse as a practical web-engineering system for secure, scalable, and governed research data collection.

Blogs remain a valuable source for understanding long-form narratives, stance formation, and discourse evolution during politically and socially salient events. However, analyst workflows still suffer from two persistent problems: broad trackers often contain large amounts of noise, and single-tracker views make it difficult to compare discourse across related but distinct corpora. This paper presents BlogTracker 2.0 as a relevance-aware and comparative analytics platform for multilingual blog discourse. Relevance-aware analysis allows analysts to define the context of interest for a tracker, score posts for relevance using an LLM, and rerun dashboard and deep-analysis views on the filtered subset so the system surfaces signal rather than corpus noise. Cross-tracker comparison provides side-by-side analysis of two trackers across coverage and overlap, content signals (sentiment, toxicity, grievance), shared semantics (entities, narratives, clusters), influence, language mix, and network structure. We demonstrate both workflows using the Indonesia protest tracker (105~blogs, 152,235~posts) and a companion Nepal protest tracker (112~blogs, 63,831~posts), showing how the same enrichment stack supports focused single-tracker investigation and comparative multi-tracker sensemaking.

This paper presents VTracker 2.0, a publicly available video analytics platform for monitoring YouTube content at scale. The system supports tracker creation, automated ingestion, background analysis, and interactive dashboards. We highlight two capabilities aimed at protest and crisis monitoring: symbol analysis, which scores video keyframes along four SCEP dimensions (social, cultural, economic, and political) and rolls results up to tracker summaries, dominant symbol ratios, and filterable channel and video lists and the tracker timeline bar on the manage page, which plots videos by publish date with category coloured ticks and interactive filtering. Using the Indonesia Protests 2025 tracker (5,966 videos collected with protest related keywords), we examine whether symbolic content aligns with viewer engagement, which SCEP dimension dominates protest related content, and whether high performing videos can succeed without visually detectable symbols. The case study points to a strongly political symbolic profile, while engagement patterns are mixed: politically dominant videos lead among symbol classified content, yet videos without detected symbols average higher views overall. We conclude that SCEP analysis complements, rather than replaces, text and engagement metrics, and outline cross platform extension as future work.

Money laundering is often expressed through the structure of a transaction network rather than through any individual record, making graph anomaly detection a natural fit for this problem. Existing tools, however, impose a practical trade-off: accurate graph detectors often behave as black boxes that investigators cannot readily interpret, while interactive visual-analytics systems rarely score, explain, or adapt autonomously. We present GlassBox-AML, an interactive system that closes this gap by detecting, explaining, and allowing analysts to correct money-laundering structures in a transaction graph, all within a lightweight, browser-only application that requires no backend. As transactions arrive, an unsupervised structural detector scores every account, and suspicious substructures are highlighted in a force-directed graph, heat-colored by risk and annotated with automatic typology badges that name patterns such as smurfing fan-out, fan-in consolidation, layering cycles, and mule rings. A single click on an alert reveals a glass-box feature breakdown and a model-faithful counterfactual that identifies the single transfer whose removal most reduces the risk; a confirm-or-dismiss action then adapts the detector online. On five bundled datasets, the detector achieves ROC-AUC 0.98--1.00 and PR-AUC 0.88$--$1.00, and maintains ROC-AUC near 0.99 down to a 1.4\% illicit ratio, combining strong detection with an interface designed for fast and confident analyst decisions.

Legal large language models (LLMs) increasingly support reasoning across jurisdictions, courts, and legal systems, yet legal validity is fundamentally bounded by jurisdiction-specific authorities. Existing legal LLMs often transfer semantic patterns without preserving authority scope, leading to rule mixing, wrong-jurisdiction grounding, and conflict masking across legal systems. We formulate this problem as authority-bounded cross-jurisdiction legal reasoning. To address it, we propose a verifier-centered legal reasoning framework that combines jurisdiction-scoped retrieval, authority verification, and conflict-preserving multi-agent synthesis. We further introduce authority-sensitive evaluation dimensions, including citation fidelity, wrong-jurisdiction grounding, and conflict detection. Experiments across four cross-jurisdiction legal reasoning settings show that the proposed framework consistently improves authority-valid reasoning reliability over direct fine-tuning, single-agent retrieval-augmented generation, and generic multi-agent reasoning baselines. Our results suggest that reliable legal transfer requires preserving authority validity rather than transferring semantic similarity alone.

Mapping vulnerability reports to their corresponding source code repositories and vulnerable versions is essential for vulnerability analysis and software evolution. However, establishing links between Common Vulnerabilities and Exposures (CVEs) and concrete code artifacts remains challenging. Existing approaches predominantly rely on patch references or explicitly curated database links, resulting in limited coverage. In this paper, we present a framework that constructs traceable mappings from CVE identifiers to downloadable vulnerable code snapshots. It integrates heterogeneous sources to establish reliable CVE-to-repository mappings, and introduces a heuristic propagation strategy that groups CVEs by software entities (vendor, product) to infer missing repository links. Based on these mappings, we perform version localization to identify vulnerable releases and acquire code snapshots. We evaluate our approach on 253,775 CVE records from 2015 to 2025. Our framework successfully maps 73,328 CVEs to repositories (28.89% coverage), representing a 258.94% improvement over direct-link baselines. Moreover, it provides traceable, versioned code snapshots for 31,799 CVEs, creating a large-scale, representative, and actionable dataset that effectively bridges vulnerability identifiers to their exact code context.

LLM evaluation pipelines often have many candidate judges: general LLM-as-a-judge prompts, reward models, safety classifiers, confidence variants, and task-specific verifiers. The deployment question is not only which judge is best, but which judges should be called, on which examples, and when panel construction should stop. We formulate judge-panel design as a role-conditioned allocation problem. From a small labeled audit set, declared slices, and judge costs, the method estimates target-relative roles: copies add no conditional information, complements improve the global panel, and specialists help only on slices. These roles induce a policy: drop copies, add complements globally, route specialists conditionally, and stop when validation gain falls below a threshold. Across reasoning, code, safety, preference, reward-model, summarization, and math audits, the method is compared with single judges, flat panels, matched diversity heuristics, full-call stacking, reliability juries, and frugal cascades. The result is a regime map for judge calls: route specialists on deployable slices, stop in saturated verifier regimes, keep broad ensembles when their risk benefit is worth the cost, and ignore conditional copies. The output is a reusable, auditable call plan for the next evaluation batch.

LLM-as-a-judge systems are increasingly used as scalable evaluation infrastructure for open-ended model outputs, but deploying a judge panel is a finite-label decision. Additional judges may add useful information, yet they also expand the calibration and validation problem that human labels must support. We study this tradeoff as finite-calibration panel selection (FCPS): given a target evaluation distribution, a pool of candidate judges, and a limited labeled audit set, select a judge path, deployed prefix size, and aggregation family. The resulting regime map distinguishes low-dimensional reliability/stacking aggregators from unrestricted joint-output tables. Across RewardBench, LLMBar, SummEval, and Arena100K with a seven-judge pool, validation-selected scalar or reliability aggregation is usually more stable than unrestricted joint-table calibration under the available labels. Controlled interaction regimes show the complementary case: when higher-order judge interactions are present and calibration support is sufficient, table-like aggregators become valuable. The practical output is a label-accounting protocol and diagnostics for deciding when panel complexity is supported by finite calibration data.

Clustering large-scale multilingual text remains challenging, particularly for Large Language Models (LLMs) constrained by limited context windows. We propose a structured LLM-based clustering pipeline that overcomes these limitations through chunking, local clustering, and cluster merging. We evaluate the approach on the Conference on the Future of Europe dataset, a real-world multilingual corpus of 8,000 citizen proposals, targeting both macro- and micro-categories. We assess cluster quality with silhouette, V-measure, F1, and LLM-as-judge evaluations validated against human annotations ($\kappa=0.785$). Results show the clear quality-coverage tradeoffs of varying chunk size, overlap and embedding-based filtering. Our structured pipeline outperforms a diagnostic agentic baseline that converges on classical TF-IDF clustering and achieves comparable V-measure to prior semantic-community-based approaches while maintaining full corpus coverage, ensuring that all the citizens' proposals are considered. While macro-category clustering is robust, micro-category discovery remains limited by the short textual context of proposal titles.

Scientific workflows increasingly compose independently published data services and analytical tools, yet output-type compatibility and flat evidence labels do not show that selected branches jointly satisfy a scientific evidence policy. We study the post-encoding problem: inputs are a normalized goal, a capability registry, and an executable policy. We introduce context-bound workflow DAGs with independently checkable evidence-policy certificates. A compact policy language expresses evidence existence, common ancestry, goal-context equality, direct or transitive consumption, field equality or inequality, conjunction, and disjunction. Before planning or verification, a registry validator enforces unique capability identifiers, supported templates and evidence kinds, finite nonnegative costs, provenance completeness, and stage acyclicity. Certificates bind the goal, registry, policy, canonical DAG, capability snapshots, union cost, and witness assignment. For finite validated stage-acyclic registries, exhaustive canonical enumeration is complete, and selecting the least-cost verifier-accepted root is globally optimal under union-of-capability cost, including shared subworkflows. A reproducible Python artifact evaluates 24 biomedical goals, 75 controlled capabilities, and four executable policies. Type-only reachability produced no accepted certificates. Flat evidence coverage reached 100% for every goal but passed policy checking on only 12 of 24; policy-constrained synthesis passed all 24. The checker rejected 264 certificate mutations and seven malformed registry or policy environments, including negative costs, duplicate identifiers, unknown policies, stage violations, and erased root context. A disclosed stress test shows exponential candidate growth. The guarantee is narrow: acceptance proves conformance relative to validated contracts and policy semantics, not biological truth, live-service correctness, or natural-language understanding.

Web-based LLM agents and intelligent conversational services require long-term memory to support personalized, evidence-grounded interactions across many sessions. Existing memory systems usually make a trade-off between efficiency and structure. Atomic-fact memories are easy to retrieve, but they may detach facts from the dialogue context in which they were established. Graph-based memories can represent richer entity-event relations, but they often introduce high construction, maintenance, and retrieval overhead. This paper proposes E2-Mem, an episodic-event hierarchical memory framework for long-term LLM agents. The key idea is to retrieve local conversational context before selecting fine-grained answer evidence. E2-Mem organizes dialogue history into two complementary levels: episodic memories preserve the topic, participants, temporal scope, and local context of dialogue segments, while event memories store precise answer-supporting facts under their parent episodes. At query time, E2-Mem performs hybrid lexical and semantic recall over episodic memories, expands selected episodes with child-event evidence, and reranks the results for compact answer-time memory construction. Experiments on the LoCoMo benchmark show that E^2-Mem improves evidence localization and answer quality. Compared with the Mem0 top-30 baseline, it improves F1 from 0.4224 to 0.4767 and BLEU-1 from 0.3429 to 0.3941. Its retrieval is slower than lightweight vector retrieval, but substantially faster than graph-database memory retrieval while keeping the prompt memory budget comparable to Mem0. These results suggest that a lightweight episodic-event hierarchy offers a practical memory design for web-based LLM agents and retrieval-augmented conversational applications.

As machine learning increasingly depends on labeled data, model-in-the-loop annotation has become the standard workflow for producing perception data in autonomous driving. But, when AI suggestions are wrong, annotators often accept them due to anchoring and automation bias, turning errors into human-confirmed labels regardless of confidence. We name this type of noise Suggestion-Conditioned Label Noise (SCLN). To study this problem, we first build AILA-Bench, a benchmark for 2D object detection: we recruit 370 annotators to label 100,000 images from BDD100K under three conditions, Human-only, AI-assisted, and AI-assisted+confidence, and we record fine-grained interaction traces and AI error types. To address this problem, we formalize the counterfactual question of whether an annotator would still make the same error had they not seen the AI suggestion as a risk score SCLNScore, and we give the assumptions under which this score is identifiable. We then combine the annotations with ground truth to train two probabilistic models, p1 and p0, with XGBoost to estimate this difference. Experiments show that AI assistance introduces a different error structure and only gives a limited reduction in overall error rate. Compared with Human-only annotation, whose FDER is 31.5%, AI-assisted annotation with confidence decreases FDER to 30.8%. The main change is in the distribution of errors: localization errors rise from 3.9% in Human-only annotation to 13.9% under AI-assisted annotation. Among 567,963 source-linked final labels, 95,601 are linked to wrong AI suggestions, and 72.5% inherit the same error type. SCLNScore detects these AI-induced errors more effectively than confidence-based review, reaching AUPRC 0.472 and a 2.322x gain at a 1% review budget; after transfer to nuImages, it remains above random with AUPRC 0.418 versus a 0.308 baseline. Demo: https://aila-bench.github.io/AILA-Bench/.

Tabular reasoning is central to decision-support and financial AI applications, requiring models to reason over structured evidence through schema understanding, value selection, computation, and logical comparison. External tools can make tabular reasoning more explicit by converting implicit operations into executable intermediate steps. However, existing prompting and fine-tuning approaches rarely model tool use directly. This makes it difficult to learn effective tool-use policies over multi-step trajectories. We introduce TabularTool, a tool-augmented framework that formulates tabular reasoning as a sequential decision-making problem with a large language model (LLM) controller. At each step, the controller first decides whether a tool is needed. When necessary, it selects and parameterizes a tool from a compact portfolio, forming an ordered tool-use trajectory. To support policy learning, we construct a curriculum-guided trace dataset. We also design a decomposed reward function that provides dense feedback on format compliance, answer correctness, and tool-call quality. The policy is optimized in two stages: supervised fine-tuning provides a stable initialization, and reinforcement learning further refines tool-use strategies. Experiments on three benchmarks show that TabularTool achieves the highest average performance among methods with full benchmark coverage. It also outperforms methods with partial benchmark coverage on their commonly evaluated datasets. Further analysis shows that the learned policy produces effective tool-use trajectories and that larger models achieve higher tool-assisted accuracy.

Weak signals are early, low-visibility indicators that precede significant changes before those changes become established. Existing detection methods, based on keyword frequency, topic modeling, or untyped graph topology, fail to capture the semantic and relational structure through which such signals manifest. In this paper, we propose C-Unseen, a self-interpretable framework for weak signal detection in Dynamic Temporal Knowledge Graphs (DTKGs). We define a weak signal as a rare, semantically coherent subgraph that proliferates across consecutive TKG snapshots. The framework operates through two modules: a Rare Subgraphs Extractor, in which an LLM identifies subgraphs whose content is in tension with the dominant snapshot narrative via chain-of-thought reasoning, and a Weak Signal Alerter, in which the persistence of these rare subgraphs is tracked across time steps to isolate true weak signals. Experimental results demonstrate that C-Unseen outperforms keyword-, topic-, and graph-based baselines.

Retrieval-augmented generation (RAG) with inline citations is widely seen as a route to verifiable, trustworthy answers. Yet a \emph{correct} citation is not always a \emph{faithful} one. A model often produces an answer from its parametric memory and then attaches a supporting passage by shallow lexical matching, a behaviour known as post-rationalization, so that the citation does not reflect how the answer was actually produced. Existing reinforcement-learning (RL) RAG systems reward citation \emph{correctness} (whether the cited document supports the statement) but not citation \emph{faithfulness} (whether the model genuinely relies on it), leaving the core trust problem unaddressed. We propose \textbf{FaithGRPO}, which turns faithfulness into an optimizable objective. For each statement--citation pair we compute a \emph{counterfactual-causal} signal: how much the answer changes when the cited evidence is removed, replaced, or adversarially perturbed. We combine this signal with answer- and citation-correctness terms into a group-relative reward for GRPO. To keep perturbations clean and localizable, FaithGRPO grounds generation on \emph{structured evidence units} rather than free-form spans, and we extend the counterfactual probe to the visual modality by occluding cited image regions. We evaluate citation correctness, citation faithfulness under counterfactual probing, answer quality, robustness to conflicting contexts, and the visual extension. Across ALCE and ExpertQA, FaithGRPO reduces the post-rationalization rate by $41\%$ relative to correctness-only GRPO ($34.7\%\!\rightarrow\!20.6\%$) and raises the average causal gain from $0.27$ to $0.42$ without sacrificing citation correctness or answer quality; its visual variant cuts visual post-rationalization to $22.0\%$ and improves region grounding to $62.0$ IoU.

Bug-fixing commit (BFC) identification is critical for vulnerability tracing, defect induction via SZZ, and code review prioritization. Prior multimodal approaches combine commit messages, code diffs, and issue reports through late fusion probability stacking. However, this strategy discards fine grained cross-modal dependencies at the feature level, so it cannot reliably separate genuine repairs from superficial keyword matches. We propose TriFuse, a framework that encodes all three artifacts using a single shared UniXcoder encoder. It performs self attention across modalities at an intermediate level and includes a quality gate to control the influence of the Issue channel. In addition, we present the first controlled study that isolates issue source construction as an independent experimental variable. By fixing splits, labels, messages, and diffs while varying only the Issue field, we show that the way issue text is built materially shifts measured performance, within model modality rankings, and the observed gap between mid-level and late fusion. On the high quality TensorFlow corpus, TriFuse achieves an F1 score of 0.9851, outperforming all baselines. On the larger and noisier multi-project corpus, it matches the state-of-the-art VulCurator in repeated split mean F1 once issue source construction is harmonized, while using 67\% fewer parameters. Our findings identify issue source quality, not architectural depth, as the primary bottleneck in multimodal BFC identification.

While automated content moderation is vital for ensuring platform safety, user trust, and risk compliance, conventional classifiers and zero-shot Multimodal Large Language Models (MLLMs) perform reliably on search-driven platforms like YouTube but fail on modern social ecosystems like Rednote. By relying on feature discrepancies, these standard approaches mislabel benign social rhetoric including irony and twist humor as deceptive lures, thereby triggering a severe over-moderation crisis and conservative under-detection. To resolve this dilemma, we propose SP-MOD, a training-free moderation framework that shifts the paradigm from literal semantic discrepancy tracking to socio-pragmatic decomposed reasoning. Operating as an interpretable runtime layer, SP-MOD utilizes a Stage-1 lightweight risk router with dynamic threshold gating to filter clear cases, followed by a Stage-2 case-retrieval augmented reasoning engine that decodes communicative intentions across three independent linguistic modules, concluding with an asymmetric adjudication process to block genuine traffic-seeking deception while explicitly protecting creative expressions. Evaluation on a newly curated long-tailed benchmark demonstrates that SP-MOD substantially outperforms standard baselines and strikes a superior balance between compliance enforcement and expressive flexibility.

AI-assisted security workflows are moving large language model (LLM) outputs from analyst-facing prose into executable detection artifacts: Sigma-style rules, backend queries, and deployment candidates. We study deployment validation for generated Web detection rules: a rule may pass ordinary acceptance gates, including seed replay and benign screening, yet miss benchmark-preserving, oracle-retained variants of the behavior it was meant to detect. RuleReplay is an execution-based validation methodology that records seed reproduction, natural held-out coverage, paired parent-to-transform misses, bounded background hit-rate stress, and backend-native query traces. On a scoped public Web/HTTP benchmark covering SQL injection, path traversal, and cross-site scripting, paired replay shows that, conditioned on a generated rule matching the natural parent request, the same rule misses the transformed child in 53.7% of eligible encoding/ignored-character rule-parent-child rows (95% cluster confidence interval [51.0%, 56.8%]). The effect persists under backend replay: 245 offline-accepted rules are converted and executed under three declared OpenSearch replay profiles, and the primary RawKeyword profile preserves positive transform-minus-natural gaps across four hosted generation settings. In bounded background stress, 163 of 185 RawKeyword-accepted rules make zero hits on 912 benign/background records. Prompt isolation, LLM-as-judge review, normalization, and OWASP Core Rule Set paranoia-level-1 (CRS PL1) replay reduce or expose parts of the gap, but none provides the same behavioral and backend-conditioned evidence. These results motivate runtime governance for AI-driven security operations: generated detection rules require runtime evidence, not textual plausibility alone, before they are trusted.

The rapid integration of artificial intelligence into large-scale digital infrastructures is transforming modern cybersecurity. AI-driven Cyber Threat Intelligence systems continuously process operational threat reports and classify them into MITRE ATT&CK techniques to support downstream cybersecurity operations. As these systems become autonomous, the attack surface extends beyond the AI model to the runtime data pipeline. Existing defenses, however, focus primarily on protecting the classifier and implicitly assume that incoming threat intelligence is trustworthy. This leaves a critical security gap where attackers can poison operational threat reports before inference and compromise security decisions. To address this challenge, we introduce a Runtime Governance Layer that governs the operational CTI stream before AI inference. The RGL transforms the runtime pipeline from passive preprocessing into a MAPE-K-inspired feedback loop for runtime monitoring, risk assessment, trustworthy adaptation, and data integrity verification. We formalize data poisoning as an ε-contamination process and establish a runtime governance model with provable guarantees on residual contamination reaching the classifier. The framework provides runtime assurance through a controllability invariant and operates in a single pass with constant memory overhead. Experiments on the AnnoCTR benchmark demonstrate effective poisoning detection, clear trust-score separation between clean and poisoned records, and residual contamination consistent with the theoretical bound. The Runtime Governance Layer recovers up to 73% of downstream robustness under a realistic mixed attack, while the adaptive controller maintains the safety invariant across all contamination levels with constant per-record overhead. These results demonstrate runtime governance as an effective approach for protecting AI systems at the data layer.

Chatbots are increasingly deployed in stateful workflows where conversations span multiple sessions maintain persistent conversation state. Protecting such systems requires more than confidential inference: conversation objects must be securely created, processed, stored, resumed, and isolated throughout their lifecycle. This paper presents $CChatbot$, a TEE-backed confidential chatbot architecture that models conversations as persistent security-sensitive objects governed by an explicit lifecycle. $CChatbot$ combines confidential inference with secure lifecycle management, including ownership enforcement, protected persistence, integrity validation and controlled resumption. We implement $CChatbot$ on a platform an AMD SEV-SNP confidential virtual machine and an NVIDIA H100 confidential GPU. Our evaluation shows that lifecycle management adds only 1.7\% overhead over confidential inference while secure create, store and resume operations remain lightweight. On a single-node deployment, $CChatbot$ supports up to 512 concurrent users with affordable interaction latency.

Spam detection is at the crossroads of cybersecurity and text mining, yet it has long suffered from accuracy-efficiency trade-off driven by real-time demands and class skew.Lightweight detection models are fast and resource-efficient but struggle in complex semantics due to shallow feature extraction. Large language models (LLMs) excel at deep, generalized malicious pattern recognition, yet their prohibitive latency and computational overhead fail to meet stringent real-time detection requirements. To address this issue, we propose dual spam detection (DualSpamDet), a slow-fast dual-system inference framework that enables training-free domain adaptation for lightweight models via confidence-gated LLM feature enhancement. DualSpamDet comprises three tightly components. Specifically, The Fast-System with a lightweight hybrid Transformer–Mamba backbone classifies routine emails in a single pass. The Slow-System is a frozen LLM that extracts structured semantic features for ambiguous samples. A Cognitive Collaborative Module routes samples via a confidence-threshold gating mechanism. Moreover, semantic reasoning is decoupled from classification. Fast-System makes every final decision, while Slow-System is activated only for hard cases to provide semantic augmentation. Experiments on public benchmarks (Ling, SMS, SpamAssassin, Enron) show DualSpamDet significantly outperforms most conventional machine learning methods and baseline models across inference metrics. Trained on Enron, it achieves a 14.3% recall gain on cross-domain test set and 6.5% on in-domain set, and surpasses representative few-shot and online learning methods in both scenarios, confirming its high precision and robust generalization.

Academic social networks (ASNs) rely on ranked feeds to determine which scholarly posts are more likely to be seen. Previous research has examined how scholars and scholarly content gain visibility on ASNs and how ranked feeds allocate exposure, but it remains unclear whether ranked exposure gives high-quality scholarly posts greater opportunities to be seen. In this paper, we analyze exposure allocation on SCHOLAT, a Chinese academic social networking platform. We use large language models (LLMs) to evaluate the scholarly quality of 28,392 posts and compare their evaluations with human ratings to test whether LLM-based quality assessment is reliable. We then measure each author's existing visibility on the platform and rank posts by predicted repost probability. We further use feature ablation, matched comparisons, and reranking analysis to examine whether limited exposure is more closely related to post quality or to information about authors and their platform connections. Our analysis yields four main findings: (1) LLMs can reliably evaluate scholarly post quality; (2) high-quality posts written by authors with limited visibility receive fewer opportunities to be seen under the predicted repost ranking; (3) feature ablation and matched comparisons indicate that repost prediction is more closely associated with information about authors and their platform connections than with measured post quality; and (4) reranking analysis shows that the exposure of these posts can be improved with only a small loss in predicted repost performance. These findings suggest that ranking systems in ASNs should take scholarly content quality into account alongside engagement prediction when allocating visibility.

Automating mobile privacy management requires capturing the highly contextual nature of user preferences. However, traditional runtime governance relies on static, one-shot permission decisions, which offer poor predictive stability because they fail to capture how users evaluate data-access consequences in practice. To bridge this gap, this paper introduces a trace-informed recommendation framework that models permission governance as a dynamic, closed-loop process. By systematically incorporating initial choices, fine-grained decision-process logs, and post-use consequence feedback, the framework shifts privacy modeling from static rule matching to dynamic preference inference. We evaluated our approach through a controlled user study with 72 participants, analyzing three recommendation models with incremental data richness. The empirical results reveal that users systematically revise 26.2% of their initial choices after experiencing actual service outcomes, underscoring the volatility of one-shot authorizations. Crucially, while integrating rich behavioral traces and context similarity yields only modest gains in absolute prediction accuracy, it drastically enhances user-perceived reasoning quality, system intelligence, and willingness to accept automated recommendations. This striking decoupling between algorithmic accuracy and user trust suggests a shift in how automated privacy assistants should be designed and evaluated: prioritizing verifiable, trace-based explanations over solitary prediction optimization.

To prevent harmful responses to malicious requests, LLMs are commonly safety-aligned. However, overly conservative alignment can falsely refuse pseudo-malicious instructions that are benign but contain sensitive cues. Existing mitigation methods rely on prompt rewriting, additional alignment data, or single refusal-direction interventions, making it difficult to distinguish harmfulness perception from refusal execution. To address this issue, we propose DANSE, a training-free method for hierarchical activation steering with disentangled null-space editing for over-refusal mitigation. DANSE uses layer-token-level Silhouette analysis to identify representation regions associated with harmfulness perception and refusal execution. It extracts the corresponding activation directions and learns null-space-constrained editing matrices, reducing pseudo-malicious misclassification at harmfulness perception layers while preserving refusal capability at refusal execution layers. Experiments on Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct show that DANSE achieves higher average compliance rates on over-refusal benchmarks and maintains more effective average defense rates on jailbreak benchmarks than its competitors. The results suggest that mitigating over-refusal requires fine-grained control over internal safety mechanisms, rather than simply reducing the model's overall refusal tendency.

With the rapid development of large language models (LLMs), retrieval-augmented generation (RAG) has become a key paradigm for grounding model outputs in external knowledge. However, large-scale vector databases often return redundant retrieval results, which reduces information diversity and degrades generation quality. When such retrieval is outsourced to untrusted cloud servers, sensitive embeddings, queries, and access patterns may also be exposed. In this paper, we propose an HNSW-based Privacy-preserving Reverse Skyline Query (HPRSQ) scheme, to enable efficient and privacy-aware coordination in RAG. We first design the HNSW-based Global Skyline computation and verification methods, which construct an HNSW graph over the data vectors and employ pairwise distance comparisons to guide greedy navigation and an approximate ordered traversal, followed by same-quadrant verification and global dominance computation to identify candidate vectors, and then verify each candidate through hierarchical HNSW traversal by examining at most $k_{\mathrm{ver}}$ distinct data vectors to detect potential dominators without exhaustive full-dataset verification. We then integrate these methods with the EASPE encryption framework to design HPRSQ, where pairwise relation vectors are constructed and encrypted to protect the underlying data vectors and query requests during query processing. Security analysis demonstrates that HPRSQ achieves IND-SCPA security under the defined leakage profile, and extensive experiments on three real-world datasets confirm its substantial efficiency gains over existing schemes.

The detection of hallucinations in Large Language Models (LLMs) is one of the primary bottlenecks to their reliable usage, with current methods either failing at detecting all forms of hallucination through relying on isolated signals, being uncertain but not interpretable by virtue of having black box un explainable architectures or suffering from expensive sampling overheads. To deal with these limitations, here we present DyMHD a hybrid, explainable and lightweight framework that leverages four independent but complementary evaluation metrics; multi metric hallucination detection based on semantic similarity analysis, natural language inference (NLI)based contradiction detection as well as structural assessment exploiting keyword overlap and response length features for crossmodel consistency evaluation. Features derived from these separate modules are routed into an empirically optimized Logistic Regression model that directly learns relative feature weights from labeled data and outputs a continuous reliability estimate . We create a three level decision boundary on top of the predicted continuous probability to classify responses as reliable, hallucinated, or uncertain. On the benchmark dataset of 1,500 samples, DyMHD achieves a macro F1score as high as 0.9964 and 0.9987 on Gemini and LLaMA3, respectively. Ablation studies show that although semantic similarity and structural features are the dominant drivers of most baseline predictions, contradiction detection and crossmodel consistency play critical roles as safety guardrails for state of the art level corner-case hallucinations. Fundamental empirical evidence confirms that our statistical learning engine achieves its absolute best performance as an efficient, production ready, and transparent white box hallucination detection pipeline, combining multi-dimensional semantic, logical, lexical, and structural analysis

Modern Web information systems route user tasks across heterogeneous cloud and edge nodes. Beyond the task payload, the scheduling pipeline itself continuously exposes metadata—arrival times, task types, resource footprints, node placement, and queue states—that can be correlated to re-identify users, profile behaviour, or mount inference attacks. Conventional schedulers optimise latency, throughput, and utilisation, and treat privacy as a payload-encryption concern handled outside the scheduling loop. In cloud–edge Web systems, scheduling metadata can itself become a privacy surface; placement should account for both privacy risk and efficiency. We formulate a bounded five-axis relative-risk score and use it to guide risk-aware task placement. The resulting scheduler, PASA, combines a privacy-matching score with an efficiency score through a configurable weight α. A separate catalogue maps risks to candidate protection requirements; the listed mechanisms are not executed in the present experiments. We evaluate PASA on fixed-seed synthetic workloads of up to 100,000 tasks against nine heuristic baseline configurations, with 1,000 paired Monte Carlo repetitions. On the 50,000-task Synthetic Workload A, PASA raises average privacy-matching benefit by 44.2% over the baseline mean and routes every submitted High/Crit task to a node above the synthetic capability threshold, while keeping throughput within 0.05% of the fastest baseline. Because the privacy-facing metrics reuse the risk and node signals in the scheduler, they establish internal placement-objective consistency rather than reduced success of an independent privacy attack. A high-load diagnostic further shows that the evaluated step-based dynamic-α heuristic is not deployment-ready without queue, tail-latency, and admission controls. All workloads are synthetic; no production logs or real user data are used.

Researchers organize the papers they collect into personal folder hierarchies in reference managers, and route each new paper into the folder where it belongs. This task differs from standard hierarchical text classification. A user’s folder hierarchy is not a fixed, shared taxonomy but a private and evolving folksonomy whose folder meanings may be topical, shorthand, venue-based, or process-oriented, and are often defined by the papers already stored inside them. We formalize this setting as \emph{personalized hierarchical paper routing} (PHPR): assigning an incoming paper to folders in a user-specific hierarchy without per-user training. We propose \textbf{PaperRouter-Agent}, a training-free LLM agent that grounds routing decisions in folder members rather than folder names alone. The agent first narrows the candidate hierarchy, retrieves folder-specific evidence, verifies fit by inspecting member papers, and incorporates similarity-gated feedback from past user rejections. A formative study on real personal libraries shows that PaperRouter-Agent raises overall Recall@1 from $0.39$ to $0.61$ and Recall@3 from $0.57$ to $0.83$, with the largest gains on organizational folders defined by metadata such as venue or year, where single-shot methods collapses (Recall@1 $0.09\to0.50$). On the public LaMP-2 benchmark, the same approach improves accuracy from $44.5\%$ to $51.5\%$ ($+9.0$ macro-F1) over a single-shot baseline, while remaining low-cost for practical use.

The simulation of human behavior through Large Language Model (LLM)-driven agents has emerged as a promising paradigm for studying social dynamics, decision-making, and behavioral responses at scale. However, the open challenge is how to ground agent personalities in structured behavioral representations, rather than relying on informal or unvalidated character descriptions. Without a systematic approach to personality grounding and validation, the behavioral consistency of simulated agents, and the interpretability of downstream analyses, cannot be reliably ensured. This paper presents a framework for instantiating and psychometrically validating personality-grounded LLM-agents based on the Myers-Briggs Type Indicator (MBTI). Sixteen autonomous agents are defined through behaviorally grounded trait descriptions and \textcolor{black}{assessed through an independent psychometric instrument functioning as a manipulation check for within-construct behavioral consistency}. All sixteen agents consistently reproduce their intended MBTI profile under the evaluation instrument, indicating that behavioral grounding without explicit labeling can produce stable and dimensionally coherent responses. To demonstrate the framework's applicability, a case study on phishing susceptibility simulation is presented. Phishing provides a suitable application domain because susceptibility to social engineering attacks is strongly influenced by individual psychological and behavioral differences. The simulation reveals meaningful differences in phishing susceptibility across personality profiles, supporting the framework’s ability to capture behaviorally relevant variation in agent responses. Beyond this domain, the proposed framework offers a reusable methodology for deploying personality-grounded agents in simulation contexts requiring diverse and structured behavioral representations, including social science research, user experience testing, education, and healthcare.

With the rapid expansion of cloud computing, microservices, and distributed systems, conventional intelligent operations methods that rely on rules and expert experience are increasingly unable to cope with complex and dynamic environments. Recent large language model (LLM)-driven agent technologies provide a new technical direction for intelligent operations, but existing methods still lack mechanisms for continuous experience accumulation, knowledge abstraction and reuse, and autonomous capability evolution. To address these limitations, this paper proposes a role-oriented autonomous evolution framework for intelligent operations based on an experience-driven LLM agent. The agent is defined as a digital operations entity with role responsibilities, autonomous learning capabilities, and continuous evolution capabilities. Through observational learning and autonomous exploration, the framework constructs multi-level experience knowledge and ontology knowledge, and introduces an experience-driven Observe-Retrieve-Plan-Act-Evaluate (ORPAE) closed-loop decision mechanism to coordinate experience acquisition, knowledge evolution, and task execution. The framework further integrates a unified observability data foundation, a long-term memory mechanism, ontology self-learning and reasoning, and a Model Context Protocol (MCP)-based tool ecosystem, thereby supporting environment perception, experience-enhanced reasoning, knowledge generation, and intelligent decision-making in complex distributed environments. Experimental validation shows that the proposed agent framework improves autonomous execution for complex intelligent operations tasks and provides a feasible technical approach for building continuously learning intelligent operations agents.

Observability metric stores often contain thousands of structurally similar time-series tables, where table-level semantics depend heavily on metric names, descriptions, and operational meanings rather than column names alone. This makes schema context selection a critical step for NL2SQL systems: insufficient context may miss necessary metric tables, while excessive context may introduce interference from weakly relevant metrics. This paper studies schema context selection for querying CockroachDB metrics stored in GreptimeDB. We build a query-context optimization pipeline that combines vector retrieval, Qwen-based second-stage filtering, schema reading, XiYan-based SQL generation, execution, and LLM-as-Judge evaluation. The system is evaluated on 1,500 leveled natural-language queries covering single-table, aggregation, time-range, multi-table comparison, and ambiguous diagnostic scenarios. The paper makes three main contributions. First, it designs and evaluates a two-stage schema context selection pipeline for large-scale observability metric stores. Second, it compares pure vector retrieval, semantic filtering, candidate-pool expansion, and large-context no-selection baselines, showing that selected and compact schema context improves query quality beyond simply providing more schemas. Third, it analyzes the interaction between model scale and context selection, and further identifies remaining bottlenecks in complex multi-table and diagnostic queries, including temporal alignment, metric composition, multi-table joins, and diagnostic intent planning.

Ensemble learning has been investigated as a promising strategy to improve the performance and robustness of Knowledge Graph Embedding (KGE) models. However, conventional ensemble methods require training, storing, and deploying multiple models, leading to substantial computational and storage overheads. Weight averaging methods instead exploit a single model's training trajectory to construct an implicit ensemble in parameter space, mostly with the storage cost of a single copy of the model's parameters and without additional inference overhead. In this paper, we present the first systematic evaluation of weight averaging approaches for KGE models under controlled graph noise, benchmarking a broad range of methods along with an ensemble baseline across representative KGE architectures and benchmark knowledge graphs. Our results show that weight averaging improves link prediction accuracy most of the time at every noise level we test, while its degradation rate as noise increases remains close to that of a single model. We release our experimental protocol and benchmarks to support future research on ensemble alternatives for KGE models.

Cyberbullying detection in Arabic social media remains challenging due to the scarcity of large-scale labeled resources. We propose a reaction-based distant-supervision framework using 193,030 Arabic Facebook posts, where audience reactions provide weak labels but are excluded from classifier inputs to prevent label leakage. We evaluate word-level TF-IDF, character n-gram, and transformer representations. AraBERTv2 achieves the highest Macro-F1 (0.827), followed by character n-gram LinearSVC (0.807) and word-level TF-IDF LinearSVC (0.788). Additional robustness experiments with the character model yield Macro-F1 scores of 0.798, 0.777, and 0.824 under temporal, source-disjoint, and duplicate-safe splits, respectively. These findings demonstrate the utility of reaction-based weak supervision while emphasizing that reaction-derived labels represent audience-informed proxies rather than ground-truth abuse annotations

This paper presents Lumyo, a mobile learning application that uses AI to combine personal note management with source constrained Retrieval Augmented Generation (RAG). Unlike general purpose conversational assistants that rely mainly on model knowledge, Lumyo allows students to explicitly select the notes that define the retrieval context for each conversation. Notes are divided into chunks and indexed asynchronously in a vector database, while user queries are processed through semantic retrieval restricted by both user identity and the identifiers of the selected notes. The retrieved fragments are provided to a large language model together with recent conversation history and the generated response includes references to the source notes. The system also uses AI for summarization and categorization to support note organization and review. The proposed design aims to combine the efficiency of generative AI with learner control, transparency and the use of students' own learning materials as the basis for generated responses.

Traditional learning platforms largely follow a one-size-fitsall model, adjusting only content difficulty and pace, leaving learners with high anxiety, low confidence, or weak self-regulation disproportionately likely to disengage. This paper presents ATLAS (Adaptive Tenfactor Learner-Aware System), a web-based platform modelling ten cognitive, affective, and personality dimensions, consolidated from the LAMPAS framework. A deterministic Profile Analyzer converts psychometric measurements into twelve auditable, natural-language instructions that guide a single LLM generation pass, producing lessons that are both cognitively scaffolded and affectively coherent, including rendered diagrams and interactive exercises. ATLAS was evaluated in a small-scale study of 25 students (20 distinct profiles), using a four-condition comparative ablation. Full ten-factor profiling was judged significantly better adapted than narrower profiling scopes (paired Wilcoxon tests, all p < .01, medium-to-large effects), with a larger, though exploratory, gain among learners with affective vulnerabilities that existing platforms do not serve. A human self-matching task independently corroborated this preference (75%, p = .021). These results provide measurable comparative evidence of adaptation quality and perceived personalisation (mean rating 4.5/5), though learning-gain effectiveness was not measured and remains future work.

Our study investigates training data influence and fairness for utilizing machine learning (ML) models to detect bank account fraud. We utilize the Bank Account Fraud (BAF) Dataset, which contains 1 million synthetically generated instances to train a ML model that maximizes fraud recall while maintaining a false positive rate below the 5% threshold. We computed the group leave-one-out influence values for the following features: month, customer age, income decile, name email similarity, housing status, employment status, and payment type. We found that each individual feature group did not significantly impact the model’s performance. However, when we increased the group size for each feature, we found that the magnitude of the influence estimation increased. Additionally, we calculated the predictive equality for the three protected attributes in the dataset: customer age, income decile, and employment status. Using our predictive equality calculations, we identified groups within each feature with relatively high false positive rates and rendered them influential in our fair ML evaluation. Our results reaffirm the importance of taking careful considerations when utilizing ML for impactful decisions such as bank account fraud. We recommend applying a value sensitive algorithm design to best understand the impact and costs of optimizing fraud recall, maintaining low false positive rates, understanding training data influence, and evaluating fair ML outcomes.

Corruption is among the most persistent societal challenges confronting governments, organizations, and individuals worldwide. Information systems, including decision-making algorithms based on artificial intelligence (AI), are often proposed as a remedy, because of their potential to offer more standardized and objective alternatives to human discretion. However, for AI to help, people must accept the decisions it produces, and existing research shows that people are often averse to algorithmic decision-makers, particularly for evaluative tasks such as hiring. We argue that perceived corruption of the context changes the meaning of human judgment: where the context appears clean, human discretion reads as a positive aspect; where it appears corrupt, that same discretion is reinterpreted as a drawback. Perceived corruption should therefore weaken algorithm aversion, not by improving perceptions of AI, but by reducing the perceived advantage of the human alternative. In a vignette experiment with 917 U.S. adults, we manipulated the decision-maker (hiring manager vs. AI system) and corruption in a hiring context (clean, moderate, high). In the clean context, the hiring manager was rated 1.90 points fairer than the AI system; under corruption, that advantage reversed, and participants assigned to the hiring manager shifted from preferring a human reviewer to preferring the AI. To understand why, we measured five perceived properties of the decision process: discretion, authority, subjectivity, quantification, and decontextualization. Quantification and decontextualization accounted for the algorithm's disadvantage in a clean context and for much of its reversal under corruption: the corrupt manager came to seem nearly as reductive as the algorithm, so these properties no longer separated the two decision-makers in the corrupt condition.

Heterogeneous data in Web-oriented intelligent analysis often contain dependencies across sources, scales, and semantic attributes that cannot be adequately represented by conventional flat-vector learning. This paper proposed HMGF-MI, a metadata-guided heterogeneous feature graph classification algorithm based on relation-aware learning. The algorithm first constructs a multi-band and multi-window candidate feature space and associates each descriptor with metadata describing its feature domain, frequency band, temporal window, class relation, and spatial pattern. A spatial-functional channel graph represents source-level dependencies, while candidate descriptors are transformed into nodes of a heterogeneous feature graph whose weighted edges encode metadata consistency and structural similarity. Relation-aware feature scoring then integrates mutual-information relevance with graph-neighborhood support, allowing discriminative evidence to propagate across structurally related nodes. A redundancy-controlled selection strategy further suppresses highly similar descriptors and produces a compact feature subset. The selected graph descriptors are processed by lightweight classifiers, and their posterior probabilities are fused for robust classification without relying on high-capacity models. Motor-imagery EEG is used as a representative evaluation domain containing noisy, heterogeneous, and structurally related descriptors. On BCI Competition IV-2a, HMGF-MI achieves 83.40% accuracy, 83.19% macro-F1, and 76.85% Cohen's kappa. Additional experiments on BCI Competition IIIa and BCI Competition IVa demonstrate its applicability across different channel scales and class configurations.

Financial misinformation detection faces significant hurdles due to its semantic subtlety, knowledge dynamism, and rhetorical manipulation. End‑to‑end classifiers struggle with the hidden context dependencies and cross‑temporal contradictions inherent in financial discourse. Existing Retrieval‑Augmented Generation (RAG) and tool‑augmented methods rely on static retrieval or fixed agentic pipelines that lack adaptability to diverse deception patterns. In this paper, we propose CATO (Context‑Aware Tool Orchestrator), a novel framework that dynamically composes analytical pathways via a metacognitive planner. CATO constructs an Evidence Propagation Directed Acyclic Graph (DAG), enabling multi‑granular verification by orchestrating a specialized suite of 9 tools—ranging from fact‑anchoring to global reasoning. By leveraging Dempster‑Shafer theory for evidence fusion, CATO integrates multi‑source signals into an uncertainty‑aware veracity score. Experimental results on the FinFact and FinGuard benchmarks show that CATO out‑performs strong LLM, RAG, and agent baselines, achieving a 3.90% F1 improvement over the strongest FinFact baseline while providing human‑auditable explanations for its reasoning process. Our code is available at https://anonymous.4open.science/r/CATO‑4FE6.

Artificial intelligence (AI) has become a fundamental component of the operational landscape. It relies on an intensive knowledge architecture that combines several aspects, including large language models (LLMs), generative AI, Agentic AI, Retrieval mechanisms, and domain-specific knowledge models, making it vulnerable to several attacks. Existing research primarily focuses on attacks against user inputs, training datasets, and model behavior, neglecting another core parameter of an AI system: knowledge. This paper proposes a conceptual framework that defines and analyses poisoning of knowledge structure that intercepts how an AI system processes, calculates, and generates. The framework distinguishes knowledge model poisoning from traditional data poisoning and prompt injection, in which knowledge plays a vital intermediary role among raw data, model training, outputs, and decision-making.

Large Language Models (LLMs) often produce incomplete or unsupported answers in context-intensive question answering, where reasoning depends jointly on the question and the source context. Existing automated Chain-of-Thought (CoT) methods rely mainly on question-level similarity for exemplar selection, which limits their ability to capture context-dependent reasoning paths. In this paper, we propose Scholat-CoT, an evidence-grounded CoT framework for context-intensive question answering. Specifically, Scholat-CoT first builds a high-quality CoT candidate pool through evidence-grounded reasoning and consistency-based integration, and then introduces Holistic-Context Maximal Marginal Relevance (HC-MMR) to select diverse and context-aligned exemplars based on the joint representation of the question and source context. We construct Scholat-QA, a scholar-centric question answering dataset, and conduct extensive experiments across multiple LLMs. Experimental results show that Scholat-CoT achieves the highest F1-Score across all evaluated LLMs on Scholat-QA and remains competitive on general reasoning benchmarks, including AQua, CSQA, and Coin Flip.

Equipping compact small language models (SLMs) with specialized scientific expertise traditionally requires manual dataset curation by domain experts—a bottleneck severely constrained by high cognitive workloads, high costs, and limited scalability. While knowledge distillation from frontier large language models (LLMs) offers an efficient alternative for transferring specialized capabilities, standard black-box distillation pipelines frequently propagate stochastic teacher hallucinations into downstream student weights. To resolve this structural vulnerability, we introduce a reference-anchored knowledge injection framework driven by GraphRAG-gated multi-teacher distillation. Our architecture leverages a heterogeneous multi-teacher ensemble of frontier models to maximize generative data diversity. This synthesis pipeline is governed by a GraphRAG quality gate that extracts a localized domain knowledge graph directly from authoritative source literature to construct high-fidelity reference anchors, systematically pruning ungrounded or contradictory teacher responses via a quantile-based verification strategy. Validated within the conceptually dense vertical of mesoscopic physics, our framework yields substantial student performance gains across complex, multi-hop scientific queries, establishing robust token-level precision and semantic alignment. Finally, we demonstrate that pairing multi-teacher generative diversity with strict GraphRAG verification successfully constructs highly precise domain specialists from scarce text resources.

Skeleton-based action recognition on skeleton key points is important for early potentially violent prediction in public video surveillance. However, occlusions and crowded environments seriously impact the recognition accuracy, especially before a violent action is fully performed. Therefore, to address these challenges, we propose an Emotion-driven Action Recognition (EAR) model that integrates skeleton action features with emotion information for action intention prediction. The EAR model analyzes the videos frame by frame and divides each video into precursor action, middle action, and postcursor action stages, enabling potential actions to be predicted from partial observations. Experiments on public surveillance and crowd-video datasets show that the proposed method achieves prediction accuracies of 59.87% using only the precursor stage and 76.78% using the precursor+middle stages. It also achieves an AUC of 0.894, a recognition accuracy of 93.26% for fighting, and an average accuracy of 90.83% across 11 high-risk actions. These results demonstrate the potential of emotion-driven action modeling for early action Recognition and AI-assisted public video surveillance in complex surveillance environments.

Universities publish large volumes of academic information covering courses, learning outcomes, competences, assessment systems, and teaching activities. Although this information is publicly available, it is commonly exposed through institution-specific Web platforms and human-oriented documents, limiting its interoperability, reuse, and automated analysis. This paper presents GUIA, a Web and semantic infrastructure for transforming institutional academic information into structured and machine-accessible knowledge. The architecture combines automated data acquisition, a CKAN-based catalogue, an ontology-driven transformation process, an online RDF Knowledge Graph, and downstream services for semantic search, question answering, and academic analysis. The current deployment at the target institution comprises 6,477 Course Guide datasets referencing 27,571 teaching-guide documents across the available academic years, together with a Knowledge Graph containing more than three million RDF triples and over 600,000 typed resources. Beyond document publication, the resulting semantic representation enables structured queries over academic information and provides a reusable foundation for AI-assisted services supporting transparency, academic quality processes, and access to institutional information. The results demonstrate the feasibility of deploying a semantic academic information infrastructure at institutional scale, while also identifying limitations related to data modelling, interoperability, and the evaluation of downstream AI services.

Psychosocial risks damage teachers' health and professional effectiveness, and teaching is consistently identified as one of the most stressful professions. Early-warning systems are only actionable if the practitioner receiving an alert can see why it was raised. Opaque classifiers cannot supply that account, and knowledge bases that merely store assertions cannot either: retrieving a stored triple is not an explanation. This paper presents TPSCO, an OWL knowledge base for the psychosocial context of teachers, designed so that every alert carries a justification chain that is entailed rather than asserted. We propose a three-layer separation of concerns — a description-logic terminology, a DL-safe rule layer, and an application layer — and state which reasoning task belongs to which layer and why. We reorient the property set so that the chain from a self-reported manifestation, through the burnout dimension it indicates, to a psychosocial risk factor is expressed by complex role inclusion axioms. And we externalize the alert policy as configuration individuals rather than constants embedded in axioms, so thresholds can be recalibrated without re-engineering the terminology. The knowledge base is evaluated by a reasoning ablation that separates what the ontology entails from what a plain graph store would retrieve, on a cohort of 18 teacher profiles.

Foodborne diseases pose significant global concerns, highlighting the need for data-driven pathogen tracking to ensure food security. However, the manual extraction of essential foodborne pathogen assay information from scientific literature remains infeasible. We propose FoodPass-IE, a two-phase framework that automates extraction from multi-document PubMed publications. The framework first extracts isolate IDs as anchors, then links associated assay attributes to construct isolate profiles. Building on Large Language Models (LLMs), FoodPass-IE enhances conventional zero-shot prompt-based information extraction through Chain-of-Thought (CoT) reasoning and knowledge augmentation strategies, enabling in-context learning and guiding intermediate reasoning. To maintain contextual continuity across multi-part documents, we introduce reasoning-based feedback mechanisms that use explanatory CoT to propagate dynamic context. Experiments on manually annotated datasets demonstrate promising performance in both isolate ID extraction and assay attribute extraction tasks, validating the effectiveness of the proposed approaches for pathogen tracking.

Across application domains, predictive signals often sit in unstructured free text rather than structured fields, yet turning that text into useful and interpretable features is difficult. Running large language models (LLMs) over an entire corpus is costly and hard to reproduce, while end-to-end text representations can rely on surface cues that are difficult to inspect. We present an agentic workflow that takes a prediction task and a raw text corpus as input and produces an auditable feature layer. The first two agents use an LLM to derive a task-specific predictor taxonomy and weakly label a bounded text sample; routed local extractors then process the corpus, and a deterministic builder aggregates the evidence into a dynamic, longitudinal concept bottleneck. We evaluate the framework on medication discontinuation in a longitudinal oncology cohort and 30-day readmission in MIMIC-IV. With gradient boosting, the longitudinal bottleneck increases area under the receiver operating characteristic curve (AUROC) over coarse concept buckets from 0.700 to 0.761 for medication discontinuation and from 0.576 to 0.609 for readmission. The proposed framework achieves predictive performance comparable to direct BioClinicalBERT prediction on both tasks while additionally providing explicit, interpretable, and traceable task-specific concepts. LLM use is confined to a bounded weak-labeling stage costing $24.00 and $23.39, respectively, compared with projected costs of $10,648 and $11,519 for exhaustive sentence-level LLM processing of the full corpora, demonstrating the substantial cost efficiency of the proposed agentic system.

Online dating discourse remains an underexplored domain for studying emotional manipulation at scale. This study draws on a corpus of 250,000 Facebook dating posts (2020-2024), with computational experiments conducted on a reproducible 50,000-post sample to examine manipulation-related discourse, temporal dynamics, and automated classification. Using an eight-category lexicon grounded in cyber dating abuse literature, 9.54\% of the analysis sample contained manipulation-related signals, with betrayal emerging as the most prevalent category. Sensitivity analysis demonstrated substantial variation across alternative labeling thresholds, while formal temporal analysis revealed a significant decline over time (OR = 0.901, 95\% CI [0.879, 0.924]). Imbalance-aware classification further assessed the predictive value of linguistic features beyond overall accuracy. These findings provide a scalable framework for characterizing manipulation-related discourse in online dating environments.

Visible-infrared multimodal fusion benefits remote sensing object detection, yet effectively integrating these heterogeneous modalities without mutual interference remains challenging. Conventional spatial-domain fusion strategies inadvertently entangle the RGB structural phase with the IR thermal amplitude, causing modality interference. To address this, we propose Fourier-Plugin, a parameter-efficient, frequency-guided asymmetric injection network. Leveraging the orthogonal frequency decomposition where RGB encodes high-frequency structural phase and IR encodes low-frequency thermal amplitude, we design the Fourier Energy Injector (FEI). This lightweight, plug-and-play module applies the Fast Fourier Transform (FFT) to decouple phase and amplitude, dynamically injecting the IR thermal amplitude into RGB features while strictly preserving the RGB structural phase, followed by an inverse transform back to the spatial domain. Deployed as an asymmetric side-branch, FEI leaves the RGB backbone unperturbed and fully inherits its pretrained weights. Experiments on the VEDAI dataset show that Fourier-Plugin achieves 73.9\% mAP@50, surpassing the RGB baseline by 6.0 percentage points, with only 1.6M additional parameters.

Critical emergency scenarios are created by road accidents, especially in underdeveloped nations with insufficient response infrastructure. Fast first aid delivery, accurate accident localization, early environmental visualization, and primary medical diagnosis without waiting for on-site responders are four issues that are still mostly unresolved. This paper proposes 6G-FODHelp-XAI, a comprehensive framework that combines a 1024-core 6G chip, a 40-80 km fiber-optic link, and an Explainable AI module to perform accident localization, primary diagnosis (including ultrasound image analysis), and triage classification identifying whether victims are in critical, severe, moderate, or minor condition. Hospitals receive results in real-time, allowing for preparedness before to victim arrival. The architecture uses CNT/Graphene nanocomposites (RCS (Radar Cross Section)= 0.00045 m²) with a 120 dB Faraday cage. The 6G chip has 0.87 ms delay for radio resource allocation, 99.99994% reliability (six nines), and 1.02 Tbit/s throughput. The AI (Artificial Intelligence) module saves 15-25 minutes of crucial hospital preparation time by achieving 92% triage accuracy and 94% internal bleeding detection.

Faced with the urgent need to guarantee the safety of au- tonomous vehicles, where every millisecond is critical, 6G resource or- chestration can no longer settle for raw performance alone; it must also provide complete transparency in its decision making processes. It is with this objective in mind that we present InOrc-6Gnet, an intelligent orches- tration architecture for V2X networks (Vehicle-to-Everything) that com- bines extreme temporal precision with unprecedented explainability. In order to overcome this difficulty, this paper presents InOrc-6Gnet, a novel architecture based on XRL (Explainable Deep Reinforcement Learning) for intelligent resource orchestration in 6G V2X networks. Thus, to guar- antee decision making transparency, the suggested architecture uses an O-RAN (Open Radio Access Network) framework that integrates a Near- RT RIC (Near-Real-Time RAN Intelligent Controller) that operates on a millisecond timescale and includes an attention mechanism monitored by Shapley values. The V2X-ISACNet ( Integrated Sensing and Com- munications Network) benchmark, which includes V2X communications, ISAC (Integrated Sensing and Communications), and network slicing for 6G environments, is used for the experimental evaluation. Experimental results show that the system achieves an average latency of 0.607 ms for URLLC (Ultra-Reliable Low Latency Communications) services, a 99.1% SLA (Service-Level Agreement) satisfaction rate, an 80.13% QoS (Qual- ity of Service) improvement (versus 78.0% for the baseline approach), support for up to 50,000 devices/km², and a 52% energy reduction for non critical slices. The explainability of the proposed solution is further supported by a 0.967 correlation between Shapley values and attention weights.

Knowledge tracing is committed to capturing students' knowledge mastery from their historical interactions. Simulating students' memory states is a promising approach to enhance both the performance and interpretability of knowledge tracing models. Memory consists of three fundamental processes: encoding, storage, and retrieval. Although numerous approaches have addressed forgetting in the storage stage, most existing studies overlook other memory processes and personalized forgetting patterns. To address this, this paper proposes memoryKT, a knowledge tracing model based on a cognition-inspired generative sequential framework. The model explicitly models memory dynamics through a three-stage process: (i) Learning the distribution of students' knowledge memory features, (ii) Reconstructing their learning experience to emulate memory replay and retrieval, while (iii) Embedding a personalized forgetting module within the temporal workflow to dynamically modulate memory storage strength. This approach jointly models the complete encoding-storage-retrieval cycle, significantly enhancing the model's sensitivity to individual differences. Extensive experiments on three public datasets demonstrate that our proposed approach achieves best or near-best performance among state-of-the-art baselines.

Legal judgment summarization aims to condense long and legally complex court documents into concise summaries. Despite the strong progress of large language models in abstractive summarization, faithfulness remains a major challenge: generated summaries often contain plausible but unsupported statements or reasoning drift, which is particularly harmful in the legal domain. Prior work has explored factual verification and consistency-based checks, but most methods focus on local factual correctness and fail to ensure that the summary preserves reasoning-level implications. To address this gap, we propose a Dual-Level Consistency Verification (DLCV) framework for faithful legal judgment summarization, integrating micro-level fact checking and macro-level reasoning checking. At the micro-level, we propose Fine-grained Claim Triple-Consistency checking (FCTC) method, which converts summary verification into closed-ended claim checking and explicitly addresses verifier instability. Unlike prior methods that rely only on single-shot factuality judgments, FCTC performs multi-view interrogation and enforces triple-consistency across label agreement, evidence stability, and semantic claim-evidence support, yielding a more robust micro-level faithfulness signal. At the macro level, we propose Outcome-Oriented Consistency checking (OOC) to ensure reasoning-level faithfulness. To address the challenge that a summary’s internal reasoning logic is implicit and difficult to verify, OOC introduces a novel outcome-oriented verification strategy that leverages the explicit judgment result as an indirect probe of whether the summary preserves the source document’s reasoning implications. To make this check reliable, OOC masks explicit outcome statements in the summary and performs constrained entailment. Extensive experiments on two datasets show that the proposed framework achieves significant improvements over all baselines.

The amount of labeled Web data for training Aspect-Based Sentiment Classification (ABSC) models is limited. In this research, we address this problem by investigating whether Large Language Model (LLM)-based Data Augmentation (DA) can improve the performance of HAABSA++, a state-of-the-art hybrid ABSC model. Using two LLMs, we obtain four multiple sentences DA models, accounting for data balance or proportionality, and, using one LLM, we obtain two single sentence DA models based on contextual or category augmentations. These augmentation methods are evaluated against the baselines: EDA-adjusted and BERT\textsubscript{prepend}. Using SemEval 2015 and 2016 datasets, we find that suggested LLM-based DA approaches consistently improve the performance of HAABSA++. In addition, the multiple sentences approach works better on small datasets, while the single sentence approach works better on large datasets. These results suggest that the proposed DA approaches can effectively address low-resource data scenarios for ABSC, thereby improving model performance in real-world settings.

LLM-based agentic Web service composition enables agents to transform user requests into executable Web API workflows and final answers. However, this autonomy introduces new reliability challenges. A workflow may complete without runtime exceptions but still violate the expected outcome, which makes outcome-level attribution difficult. Web services may also change in operational behavior or interface compatibility, which makes fixed service-selection strategies unreliable. Existing recovery methods further suffer from costly workflow repair because blind retry may miss upstream root causes, while full re-execution may waste successful service calls. This paper presents \textsc{FLARE}, an outcome-aware root-cause localization and minimal repair framework for LLM-based agentic Web service composition. For outcome-level attribution, \textsc{FLARE} detects outcome gaps from requirement checks and evidence from service observations and answer grounding, then traces them through an outcome-aware causal execution graph. Service drift is handled by recording runtime service evidence and repair feedback, which update service profiles and repair preferences. Repair cost is reduced by ranking repairable root-cause candidates in backward slices, selecting a root-cause-specific repair action and re-executing only the affected downstream slice while reusing unaffected service evidence. We evaluate \textsc{FLARE} on 2,254 Web service-composition executions with 789 annotated failures. The results show that \textsc{FLARE} improves faulty-decision localization and reaches a repair success rate of 0.811, while reducing token and latency cost relative to whole-workflow repair. These results suggest that outcome-aware localization and affected-slice repair can improve the reliability of agentic Web service composition.

Large language model agents increasingly answer multi-hop questions, verify claims, and conduct open-domain research with external evidence. However, strong search on an individual task does not ensure reliable behavior over a sequence of related tasks: agents can repeatedly choose non-authoritative sources, overlook freshness constraints, miss conflicts, or attach citations that do not support the generated claim. We formulate cross-task evidence policy learning, where an agent learns reusable procedures for evidence acquisition and validation from prior evidence trajectories without updating model parameters. We propose EvoSeek, which stores claim-level evidence events, induces scoped executable policies from successful and failed trajectories, retrieves applicable policies for later tasks, and audits their utility and scope. On HotpotQA, 2WikiMultiHopQA, FEVER, and SciFact, covering 26,947 validation tasks and 215,576 method-task records, EvoSeek improves average task score from 45.0% for standard RAG to 57.8%, citation faithfulness from 79.6% to 85.6%, and evidence recall from 78.7% to 89.4%, while reducing repeated-error rate from 51.9% to 35.1% under a matched retrieval-call budget.

Web product matching is a high-value entity matching problem in which labels are expensive, positives are sparse, and the most useful labels are often rare semantic corner cases rather than merely uncertain strings. Classical active learning methods are model-grounded but can miss Web-specific failure modes; recent LLM-based selectors can recognize semantic difficulty but may select plausible positives without considering the current student's blind spots. We propose \method, a failure-aware active learning framework with two key innovations. First, \method\ converts an LLM into an auditable failure analyst that returns a weak label, confidence, rationale type, and label informativeness, without treating the LLM output as gold supervision. Second, \method\ introduces backbone-coupled rationale reweighting, attaching this failure prior to strong active learning backbones such as BADGE and query-by-committee (QBC) rather than replacing them. On WDC Products, a 5-seed true-LLM run with 1,000 cached Qwen2.5-32B rationales shows that \method\ consistently improves label efficiency over LLM-only selection: \fameqbc\ improves AULC-F1 over ActiveLLM-style by 0.0194 and wins on all five seeds. \method\ also outperforms training on free LLM pseudo-labels alone (F1 +0.0187), confirming that targeted human labels selected by failure-aware acquisition remain valuable even when LLM weak-label accuracy is high. Final F1 remains within 0.0032 of the strongest QBC baseline. Across proxy-rationale stress tests and ablations, \method\ provides the largest gains when class imbalance is moderate and LLM weak-positive priors are most actionable, while remaining competitive with strong classical baselines in harder regimes.

Modeling the temporal evolution of fine-grained scientific knowledge is increasingly important for scholarly graph mining. Existing scholarly graphs usually organize knowledge around papers, authors, venues, citations, or keywords, which makes it difficult to represent the multiple statement-level viewpoints contained in a single paper. In this paper, we study chronologically constrained academic viewpoint classification on evolving scholarly graphs. The task is challenging because newly introduced viewpoints have imbalanced relational evidence: textual semantics are immediately observable, whereas structural links are initially sparse. In addition, viewpoint semantics are often shaped by higher-level topic context that evolves with historical viewpoint evidence. We propose HD-AVR, a hierarchical dynamic graph learning framework for new academic viewpoint classification. HD-AVR separates structural and semantic evidence at the viewpoint level and combines them through a node-wise gate. It further maintains a recurrent topic layer, where fixed topic slots summarize historical viewpoint evidence and provide reliability-aware context for current viewpoint prediction. Experiments on four ViewPoints Network domains show that HD-AVR outperforms the reported static and dynamic graph baselines under this chronological protocol.

Choosing a Small Language Model (SLM) for an offline, privacy-sensitive Web Information System (WIS) is a multi-objective problem: accuracy, latency, throughput, and memory must all be satisfied at once, which accuracy-only leaderboards cannot resolve. We present SLMBench, a single framework that evaluates SLMs under a uniform, locally served protocol and converts the resulting measurements into a statistically grounded, Pareto-based deployment recommendation. Applied to eight Natural Language Processing (NLP) benchmarks across ten SLMs (0.5 to 4B parameters), SLMBench identifies five deploymentoptimal SLMs, with the large majority of pairwise accuracy differences remaining statistically significant after correction for multiple comparisons. A reasoning-mode ablation further shows an order-of-magnitude latency cost for a comparatively small accuracy gain. SLMBench is released as a reusable system for statistically grounded SLM selection in resource-limited WIS deployments.

Emotion recognition in conversation (ERC) involves detecting the underlying emotion behind each utterance within a conversational context. Due to emotional similarity emerging under strong contextual dependencies, ERC models still struggle to effectively distinguish between similar emotions, even with the incorporation of contrastive learning (CL) loss. To overcome the aforementioned limitations, this paper provides a mathematical analysis of the inefficiency and insufficient robustness in existing CL loss. Based on the analysis, we propose targeted item elimination and sampling simplification to enhance the model’s ability to distinguish similar emotions. Specifically, we theoretically analyze the instability factors in the gradient of existing CL loss, which lead to gradients attenuation under certain sample distributions. Accordingly, we remove the obstructive terms from the loss function, which enhances robustness across varying sample distributions and leads to more disentangled distributions of utterance representations. Moreover, to mitigate the redundancy and errors of negative sampling, we systematically analyze the lower bound of the CL loss and derive a simplified sampling strategy that circumvents explicit negative pairwise selection by steering representations to the boundary. This enables more efficient and precise discrimination for representations. Experiments on four datasets show our method achieves state-of-the-art performance in ERC.

Time series subsequence matching is a fundamental task in many real-world applications. However, existing methods face scalability limitations and high pruning overhead due to handcrafted repre sentations and loose lower bounds. In this paper, we propose LISM, a learned index framework designed for efficient and exact time series sub sequence matching. LISM first partitions each data sequence into win dows using a J-sliding window strategy and projects each window into a low-dimensional representation via Piecewise Aggregate Approximation (PAA). The resulting points are then transformed into one-dimensional Z-order values, which are subsequently sorted and organized into buckets. A learned PGM-index is built upon the bucket boundaries to accurately predict the target bucket for each query window, thereby significantly re ducing the search cost. During query processing, the original Euclidean distance threshold is converted into a safe window-level search radius, guaranteeing no false dismissals. Additionally, LISM optionally employs a learned filter for bucket-level pruning when approximate retrieval with user-specified recall guarantees is acceptable. Extensive experiments on both real-world and synthetic datasets demonstrate that LISM consis tently outperforms state-of-the-art methods in query efficiency and scal ability.

ing approach for Flink to reduce resource utilization while guaranteeing multiple jobs’ SLAs. TEMT introduces Dynamic Constrained Bayesian Optimization to handle massive SLA constraints by leveraging correlations among jobs and compressing redundant ones, and Importance]>

Efficient cross-host data synchronization remains a key challenge in distributed databases. Existing systems typically propagate committed updates and synchronize replicas over the network, which introduces extra message exchanges and synchronization waits, and makes it difficult for query replicas to reflect the latest committed state in time. This paper presents CXL-Sync, a data synchronization mechanism based on CXL memory semantics. After a transaction commits, the primary partition uploads the latest committed updates to shared CXL memory, from which non-primary partitions directly pull and install the updates into local replicas. While CXL-Sync reduces network synchronization overhead through CXL, its design must account for CXL's higher latency and lower bandwidth than local DRAM, as well as replica freshness and visibility. Experiments on TPC-C and YCSB show that, compared with a network-based synchronization scheme, CXL-Sync improves throughput by about 2.2× on TPC-C and about 2× on balanced read-write YCSB, reduces P95 latency by about 12.2% on both workloads, and keeps read visibility delay at zero.

Graph anomaly detection plays a critical role in applications such as social network analysis, web fraud detection, spam and bot identification. Existing spectral methods decompose graph signals into multiple frequency bands but subsequently process them in a homogeneous manner. Specifically, they ignore the uneven distribution of anomaly-related information across the spectrum, particularly its enrichment in mid- and high-frequency bands, and overlook the heterogeneous relationships among different frequency bands. These limitations hinder the extraction of informative anomaly patterns and result in insufficiently discriminative anomaly representations. Interestingly, graph anomalies exhibit spectral characteristics analogous to off-tune and inharmonious patterns in audio signals: anomalies may induce uneven perturbations across frequency bands while disrupting their inherent relationships. Inspired by the "tuning--harmonization" paradigm in audio signal processing, we propose a Resonance Graph Anomaly Detection (RGAD) framework that progressively constructs resonance-aware anomaly representations through collaborative resonance modeling. Specifically, a Band Tuner adaptively calibrates the contributions of different frequency bands to emphasize anomaly-sensitive spectral patterns. Subsequently, a Band Harmonizer integrates multi-band information by explicitly modeling heterogeneous cross-band relationships through band integration and discrepancy identification. The resulting collaborative representation further enhances representation discriminability for anomaly detection. Extensive experiments on six benchmark datasets demonstrate that RGAD consistently outperforms existing state-of-the-art methods on both homogeneous and heterogeneous graphs, confirming the effectiveness of the proposed collaborative resonance modeling framework.

Entity Alignment (EA) aims to detect descriptions of the same real-world entities among different Knowledge Graphs (KG). Several embedding methods have been proposed to rank potentially matching entities of two KGs according to their similarity in the embedding space. However, existing EA embedding methods are challenged by the diverse levels of structural (i.e., neighborhood entities) and semantic (e.g., entity names and literal property values) heterogeneity exhibited by real-world KGs, especially when they are spanning several domains (DBpedia, Wikidata). Existing methods either focus on one of the two heterogeneity kinds depending on the context (mono- vs multilingual). To address this limitation, we propose a flexible framework called HybEA, that is a hybrid of two models, a novel attention-based factual model, co-trained with a state-of-the-art structural model. Our experimental results demonstrate that HybEA outperforms the state-of-the-art EA systems, achieving a 16% average relative improvement of Hits@1, ranging from 3.6% up to 40% in 5 monolingual datasets, with some datasets that can now be considered as solved. We also show that HybEA outperforms state-of-the-art methods in 3 multi-lingual datasets, as well as on 2 datasets that drop the unrealistic, yet widely adopted, one-to-one assumption. Overall, HybEA outperforms all (11) baseline methods in all (3) measures and in all (10) datasets evaluated, with a statistically significant difference.

Multi-domain recommendation (MDR) involves providing personalized recommendations across multiple distinct domains by leveraging shared information and user behavior patterns. However, overlapping sparsity between users and items across different domains arises when common users or items between domains are limited. This sparsity degrades the effectiveness of MDR by limiting the system's ability to learn from and leverage shared patterns and preferences. In MDR tasks for short videos, we observe that hierarchical tags alleviate interaction sparsity by providing a layered semantic framework that captures both coarse- and fine-grained content features, thereby enabling the model to establish richer multi-domain associations and improve the transfer of latent preferences between domains. Therefore, we propose a \textbf{H}ierarchical \textbf{T}ag-\textbf{E}nhanced \textbf{M}odel (HTEM) for multi-domain short video recommendation to address overlapping sparsity. Specifically, HTEM constructs a user video heterogeneous graph by leveraging hierarchical tag across domains as intermediary nodes. To effectively learn the hierarchical tag structure and capture the intricate relationships among users, short videos, and tags within graph, we introduce a tag-enhanced aggregation mechanism in hyperbolic space, which aggregates nodes to generate the final embeddings for users and short videos. Finally, extensive experiments on two real-world short video recommendation datasets demonstrate the effectiveness of our model.

With the development of the multi-video surveillance Web and advances in video analysis and emotional computing, emotion-related subconscious behavior recognition has gained increasing attention. This paper introduces an Emotion-Related Subconscious Behavior Recognition (ER-SBR) model that identifies subconscious actions by modeling correlations between body and face keypoints. The static neighbor-joining matrix is enhanced with non-associated body connections and a self-learning matrix, improving recognition performance. Based on the results, a "subconscious reaction–emotion–action" rule base is established, achieving an average recognition accuracy of 86.74%.To account for individual differences, an Individual Video Extraction based on Facial Expressions (IVE-FE) algorithm automatically extracts video segments for specific individuals. Using spatiotemporal geometric features, emotions and subconscious behaviors are annotated. Finally, the Individual Subconscious Behaviors Mining (ISBM) model combines time-domain emotional features with high- and low-frequency action characteristics to identify individual subconscious behaviors, achieving an average accuracy of 95.25% across eight individuals.

With the rapid growth of Web APIs, Mashup developers often need to select additional Web APIs that are both relevant to the Mashup requirement and complementary to the already selected Web API set. However, existing Web API recommendation methods usually focus on Mashup requirement modeling or historical invocation patterns, while insufficiently considering selected-API-aware complementarity. To address this problem, this paper proposes a complementarity-aware Web API recommendation via explicit and implicit requirement modeling framework. The framework first learns base representations of Mashups and Web APIs by integrating textual semantics and graph-based structural information. Then, the explicit branch constructs an LLM-enhanced selected-API-aware requirement text by combining Mashup information, selected Web API information, and LLM-inferred supplementary capability information, and fuses its encoded representation with the base Mashup representation for semantic matching with candidate Web APIs. Meanwhile, the implicit branch captures candidate-aware complementary context from the selected Web API set through feature interaction and attention pooling. Finally, the explicit and implicit scores are fused to rank candidate Web APIs. Experiments on a real-world Web API dataset demonstrate that the proposed framework achieves better recommendation performance than representative baseline methods. The source code is publicly available in https://github.com/IntelligentServiceLab/EIRM.

Large Language Model (LLM) serving increasingly underpins online Web services such as conversational agents, Web search, and programming assistants, where requests carry heterogeneous Service Level Objectives (SLOs) such as Time to First Token (TTFT) and Time Per Output Token (TPOT). Existing LLM serving systems prioritize maximum throughput and treat all requests uniformly, which leads to suboptimal SLO attainment. This paper introduces Scorpio, an SLO-oriented LLM serving system designed to maximize system goodput and SLO attainment for workloads with heterogeneous SLOs. Our core insight is to exploit SLO heterogeneity for adaptive scheduling across admission control, queue management, and batch selection. Scorpio features a TTFT Guard, which employs least-deadline-first reordering and rejects unattainable requests, and a TPOT Guard, which utilizes a VBS-based admission control and a novel credit-based batching mechanism. Both guards are supported by a predictive module. Evaluations demonstrate that Scorpio improves system goodput by up to 14.4x and SLO adherence by up to 46.5% under high load compared to state-of-the-art baselines.

Effective ad detection has long been a significant demand among users. Current methods, while prioritizing accuracy, often exhibit low recall in practice due to the complex loading mechanisms (e.g., increased JavaScript interactions) and multifaceted ad formats (e.g., video, image, text). In this paper, we propose ADSCAN, an integrated multi-modal framework designed for precise and fine-grained ad detection. It visualizes page execution as a directed graph and employs specialized detection modules for individual nodes (e.g., JavaScript, image) using advanced NLP and CV techniques. An enhanced semi-supervised learning approach is utilized for global detection to finalize decisions. By integrating these methods, ADSCAN comprehensively identifies ads and provides detailed insights. Experiments on 14,250 websites demonstrate that ADSCAN significantly improves ad detection recall while maintaining accuracy and precision comparable to state-of-the-art methods. Implementation and datasets are available at https://zenodo.org/records/15498752.

Multimodal fake news detection (FND) increasingly demands verdicts that are not only accurate but traceable—revealing how cross- modal evidence is combined—yet two coupled difficulties remain. First, text—image relations are heterogeneous: uniqueness, redundancy, and synergy coexist and vary from post to post, so a single global fusion rule is brittle and opaque. Second, the dominant modality shifts across instances, which static encoders and a fixed fusion pathway handle poorly.We present IMEX-FND, an interaction-aware mixture-of-experts frame- work that couples adaptive routing with explicit interaction decomposi- tion. A Multi-Modal Expert Gateway (MMEG) performs instance-wise, within-modality routing over specialized and shared experts and builds a CLIP-grounded cross-modal stream, yielding three refined, interaction- ready representations that adapt to the dominant modality of each post. An Interaction-aware Multi-Modal Expert Fusion (IMEF) module then decomposes the interactions among these streams into uniqueness, redundancy, and synergy, producing transparent, sample-wise weights via a modality-replacement training signal. The two stages form a single route-then-decompose pipeline whose routing and interaction weights are both inspectable, offering instance- and dataset-level traceability for misinformation diagnosis. Extensive experiments on the Weibo, Weibo-21, and Gossip benchmarks show that IMEX-FND achieves state-of-the-art performance, surpassing competitive baselines by 0.4—1.2% while offering superior traceability with fewer parameters.

Lateral movement (LM) is a key phase of Advanced Persistent Threats (APTs), where attackers exploit legitimate trust relationships to progressively expand control across a network. As modern enterprises increasingly rely on web-based authentication and trust management across services, these trust relationships have become attractive targets for adversaries. Due to the deliberate concealment by attackers, lateral movement is difficult to detect. Recent studies have employed graph-based methods to link entities within the network and detect lateral movement paths through anomalies in the graph structure. However, existing graph methods mainly focus on entity-level or path-level detection, making it challenging to provide precise and effective alerts for security analysts. To address this issue, we propose FLARE, a fine-grained lateral movement detection method that focuses on action-level anomalies. FLARE includes temporal multigraphs to describe network interaction actions in detail, a heterogeneous edge feature learning graph neural network (RCensNet) to learn graph structural features, and a time window strategy that aggregates spatiotemporal edge features, enabling precise action-level alerts regarding specific times and actions. Additionally, FLARE incorporates a historical sample enhancement strategy to address the class imbalance problem. We evaluate FLARE on the CERT and LANL datasets. Our experimental results show that FLARE achieves excellent performance among baseline methods, with an F1 score of 99.01% and an AUC of 99.06% on CERT, with an F1 score of 92.41% and an AUC of 99.72% on LANL.

Large-scale Web and enterprise information systems continuously generate endpoint audit logs that record process, file, and network activities. Provenance-based threat detection systems use these logs to construct causal representations of system behavior. However, these logs often contain massive repetitive background activities that inflate storage, graph construction, and model training costs. This paper studies whether such redundancy can be reduced while preserving downstream detection utility. Our analysis of DARPA E3 and OpTC audit logs reveals substantial redundancy, with frequent event patterns recurring across time and datasets. We design a fidelity-aware log reduction pipeline and evaluate two operating configurations. The full configuration combines temporary-object removal, template-equivalent benign operation merging, and causal-equivalence compression to maximize reduction, while the causal-only configuration applies causal-equivalence compression alone as a conservative fidelity-preserving setting. Across seven datasets, the full configuration reduces 1.22 billion audit events to 313.93 million events, achieving 74.28% event-row reduction and 58.15% provenance-edge reduction. The causal-only configuration reduces the logs to 392.26 million events, achieving 67.86% row reduction while changing the average ORTHRUS AUC by only -0.0034. These results show that audit-log reduction is a fidelity-aware operating space, where different configurations trade scalability for detection fidelity while preserving downstream causal utility. Causal-only preserves ORTHRUS ranking behavior with minimal AUC change, whereas Full provides stronger graph-size reduction when scalability is prioritized.

Software-Defined Networking (SDN) enables programmable traffic control for automated intrusion response, but many IDS-driven response systems directly treat IDS alerts as true security states. This assumption is unreliable under false positives and false negatives, where aggressive actions may disrupt benign services or missed alerts may delay attack mitigation. To address this problem, we propose BD-PPO, a belief-driven reinforcement learning framework for robust SDN intrusion response under uncertain IDS alerts. Instead of directly using IDS outputs as fully observed states, BD-PPO models them as noisy observations of a latent security condition. A lightweight Bayesian estimator maintains attack belief and uncertainty from historical IDS observations, and a PPO-based response agent selects among Recover, moving target defense, and Drop using a belief-augmented state. To reduce overreaction under low-confidence conditions, we further introduce uncertainty-aware soft action regularization, which penalizes destructive actions according to attack belief, uncertainty, and action destructiveness. Experiments on an emulated SDN testbed show that BD-PPO reduces wrong blocking of benign traffic while maintaining effective mitigation of malicious traffic under noisy IDS alerts. Ablation results further confirm that belief estimation and uncertainty-aware regularization are both important for balancing mitigation effectiveness and service availability.

Modern Web information systems expose large and evolving API and microservice ecosystems, making it difficult to compose services from natural-language requests. Existing LLM-based service composition methods improve intent understanding and API use, but often treat each request as an isolated episode, limiting their ability to reuse past executions, adapt to changing services, and recover from runtime failures. This paper presents REMIX, an experience-aware memory-augmented multi-agent framework for adaptive Web service composition. REMIX represents historical executions as typed graph memory, retrieves relevant experience through an index-then-expand strategy, coordinates specialized agents for planning, grounding, execution, supervision, and summarization, and uses a DQN-based controller to retain high-utility memories while suppressing redundant or stale traces. Experiments on a public multi-domain Web API benchmark and an anonymized industrial microservice dataset show that REMIX consistently improves planning accuracy, API grounding, and execution success over traditional service composition methods, single-agent LLM baselines, and multi-agent LLM baselines. On the public benchmark, REMIX achieves 81.88\% Plan.Acc and 65.50\% API.Acc, improving over the strongest average baseline by 2.9 and 3.2 percentage points, respectively.

In real-world recommender systems, sequential recommendation (SR) is widely used to enhance user experience. However, data distribution shifts and the emergence of new items often cause catastrophic forgetting and performance degradation of recommender systems. Consequently, continual sequential recommendation (CSR) has been proposed to address this issue. Nevertheless, many existing continual recommendation methods often rely on replaying historical data or updating a large number of model parameters, which may be impractical in some real-world scenarios. This paper proposes LR-CSR, a LoRA-based Replay-free Continual Learning Framework for Sequential Recommendation. LR-CSR is model-agnostic and can be readily integrated into various deep learning-based sequential recommendation models with only a small number of additional parameters. To mitigate forgetting, LR-CSR maintains a subspace spanned by historical inputs, computes its orthogonal complement to initialize and freeze the LoRA down-projection matrix, effectively suppresses interference from new tasks to old ones without requiring replay of historical inputs. Extensive experiments on three real-world datasets and two backbones validate LR-CSR's effectiveness and advantage. Code of LR-CSR is available at https://github.com/super-starch/LR-CSR-WISE2026.

Existing test-time prompt optimizers for LLM-based multi-agent systems, especially hierarchical reflection-style systems, mainly maximize reward on a small calibration set. In practice, that is not enough: a pipeline can work on the calibration examples and still break on paraphrases or mild peer-style changes. We call this failure mode interface fragility. We present BRIDGE, which turns each calibration example into a small set of stress tests for inter-agent communication. We use these stress tests to find where the handoff breaks, then keep a local repair only if it helps on held-out cases. We evaluate BRIDGE in the hierarchical reflect setting on six reasoning and coding bench marks under two worker-evaluator configurations, including a transparent rule-based gate and a Bayesian repair-gate variant. Both the rule based BRIDGE and its Bayesian repair-gate variant deliver substantial and consistent performance gains over Hierarchical MAS + MASPO across both the Qwen3-8B/Gemini-2.5-pro and the matched DeepSeek v4-flash/DeepSeek-v4-pro settings. The diagnostics show that peer-style and rewrite probes mainly serve as stress tests for identifying brittle handoffs. These results suggest that calibration data is most useful when it exposes and repairs brittle interfaces rather than serving only as a scalar reward source.

Frozen vision-language models (VLMs) are attractive for accessibility-oriented visual services because they can be deployed without task-specific retraining, but they remain unreliable in sparse expert domains. Isolated sign recognition is a representative case: signed languages encode lexical contrast through hand shape, orientation, location, contact, motion, and visible non-manual cues, so near-homonymous signs such as \textsc{think} and \textsc{know} can be confused when a VLM is queried once in an open loop. We introduce \textbf{SignAgent}, a parameter-free inference-time layer that wraps a frozen VLM with phonologically grounded exemplar memory, hybrid visual--lexical recall, dimension-wise comparison, confidence-triggered reflection, and human-audited memory consolidation. Rather than presenting reflection as a generic second thought, SignAgent uses an attribute-selective evidence-gathering rationale: when the current gloss is uncertain, the model refines only the unresolved sign dimensions and re-retrieves exemplars that better separate visually confusable candidates. We also construct \textbf{SignBench}, a diagnostic benchmark of 3,234 ASL and CSL instances across static recognition, dynamic recognition, and gloss-conditioned reasoning. Evaluation on 18 VLMs reveals human--model gaps up to 72.2 percentage points. Across four representative architectures, SignAgent improves L1 static recognition by 20.7 pp and L2 dynamic recognition by 13.7 pp on average; their mean gain is 17.2 pp. Controlled decomposition with a memory-free Blind-Reflect control shows complementary but non-additive effects: on Gemini L1, Blind-Reflect alone improves +15.5 pp, Recall+Compare alone improves +19.1 pp, and the full loop improves +27.8 pp. A Shapley-style attribution assigns 15.7 pp to memory-grounded comparison and 12.1 pp to reflection, so blind self-refinement is useful, but the best performance requires coupling it with contrastive exemplar memory.

Large language models have demonstrated strong capabilities in open-domain text generation. However, in safety-sensitive applications where helpfulness must be preserved, reliably steering model behavior remains challenging. Existing methods often struggle to adapt dynamically to input-specific semantics. To address this challenge, we propose Latent-Reasoned Prefix Steering (LRPS), an input-conditioned controller for a frozen base language model. LRPS refines the prompt representation through iterative latent-state updates, projects the final state to a compact control vector, and maps that vector to prefix key-value states that persist throughout decoding. Experimental results on Qwen2.5-0.5B-Instruct and Qwen2.5-7B-Instruct show that different latent-reasoning depths yield distinct safety–helpfulness operating points rather than a strictly monotonic change in control strength. Among the evaluated configurations, T=1 provides the highest benign-request helpfulness retention, while T=8 achieves the lowest AdvBench attack success rate and stronger unsafe-request refusal than the shallow configuration. These results suggest that iterative latent refinement provides a configurable mechanism for input-conditioned prefix steering under a frozen base-model setting. At the same time, the current implementation introduces additional computation through an auxiliary reasoner, and its scalability to larger model families remains to be further validated.

Scientific research agents require long-term memory to preserve task state across interrupted sessions, integrate heterogeneous information, and accommodate factual updates. Many memory-augmentation methods encode heterogeneous information into a monolithic vector space and rely on static semantic retrieval, so they degrade after context resets, the loss of short-term working context between sessions. Inspired by human memory, we present the NeuroStream Memory System (NSMS), which decouples input into five streams: semantic, episodic, procedural, pattern, and reference. To handle frequent factual updates, NSMS introduces a versioned storage and conflict resolution grounded in cognitive schemas, using structural constraints to reduce retrieval noise and search overhead. We construct ResearchDialogue, a scientific dialogue benchmark with explicit context resets. Experiments show that NSMS improves cross-session integration and remains competitive on historical-session retrieval, while showing advantages in single-hop factual conflict resolution on MemoryAgentBench. For contexts exceeding 150K tokens, its gated retrieval reduces computational overhead and latency.

Late-interaction visual document retrieval (VDR), exemplified by ColPali-style models, has become the dominant paradigm for retrieval-augmented generation (RAG) over visually rich documents because it indexes page \emph{images} directly and avoids lossy OCR. Its central obstacle is cost: a single page is encoded into hundreds of patch vectors, inflating the index to hundreds of kilobytes per page before ANN overhead and making nearest-neighbour search expensive. The prevailing remedy, data-agnostic sign binarization followed by full-precision rescoring, collapses at the per-token dimensionality used by late-interaction models, forcing large oversampling factors that erode the very speedup binarization is meant to provide. We argue that this is a \emph{learning-to-hash} problem rather than a quantization problem, and present \textbf{BitCol}, the first end-to-end learned binary late-interaction representation for visual documents. BitCol (i) replaces the projector with a learnable hashing head trained directly against a \emph{Hamming-MaxSim} ranking-distillation objective from a full-precision teacher, (ii) adds an \emph{evidence-preserving} regularizer that keeps answer-supporting patches locatable in Hamming space, and (iii) optionally activates \emph{query-conditional sub-codes} that select a code subspace per query intent. We evaluate BitCol on visual document retrieval, end-to-end RAG, evidence grounding, and the storage/latency Pareto frontier. BitCol retains roughly 97--99\% of full-precision ColQwen2 retrieval quality while cutting the oversampling factor from $9\times$ to $2\times$ and using a 24\,KB/page binary payload at $K=256$.

Retrieval-augmented generation (RAG) underpins many LLM web-based applications, yet deployed systems track which retrieved chunks contributed to an answer rather than which upstream sources actually caused it. This gap becomes critical under content-propagation attacks, where adversaries poison a single source whose syndicated copies dilute per-chunk detection signals, raising a clear web security and trust management concern. To close this gap, we introduce Upstream Provenance, a training-free approach that combines a single-pass Fisherinformation approximation of per-chunk causal influence with a graph- Laplacian-regularised aggregation over a content-flow graph built from near-duplicate, paraphrase, and provenance signals, lifting attribution from chunks to sources in closed form and turning the fusion of retrieved web content into a trust signal. The approach adds only one extra forward pass and is backed by a separation theorem that guarantees detection amplification and source-identification recovery under stated propagation. Across HotpotQA, MS MARCO, and Natural Questions on three open-weights LLMs, the method gains +0.14 to +0.19 detection ROCAUC over the strongest baseline at kd=4, identifies the contaminated source in 87% of injected traces, cuts residual attack rate by 2.3× over RAGForensics, and runs 13× faster than ContextCite at 74 ms per query.

Artificial intelligence voice cloning technology has achieved rapid development driven by machine learning. Although these developments have brought significant progress, they have also raised concerns about the misuse of AI voice cloning technology, causing economic losses and negative public perception. Despite the existence of numerous excellent fake audio detection systems, these systems only perform post-detection rather than pre-emptive prevention. To address this challenge, we have transformed passive measures into active ones by proposing a novel active defense method, VocalCrypt, which embeds pseudo-timbre, or jamming information, imperceptible to the human ear in audio segments based on masking effect technology, thereby creating audio fragments that cannot be voice-cloned. The masking effect is a psychological effect that reveals parts of audio that are imperceptible to human ears and is currently widely used in audio compression algorithms. Our method protects voice security without affecting voice quality. Compared to existing methods such as adversarial noise embedding, VocalCrypt significantly improves robustness and runtime efficiency, achieving a 200% increase in generation speed while maintaining interference effectiveness. Extensive experiments based on the VCTK corpus dataset demonstrate that our AI clone voice defense system performs excellently in ASV and MOS tests while maintaining the integrity of the protected audio.

Web-scale data services are increasingly deployed in serverless and cloud-native environments, where tenants are activated on demand and charged according to resource consumption. This model improves elasticity but creates a tension between memory efficiency and service-level performance: reclaiming cached state saves memory, whereas discarding it entirely may cause severe cold-start penalties when tenants return. Existing proactive resource-management methods usually operate at a coarse granularity and cannot exploit the fine-grained reuse patterns of cached data. This paper presents Arrietty, a risk-constrained cache-retention framework for SLA-aware serverless data services. Arrietty retains an adaptive subset of reusable data during idle periods and jointly optimizes the retained size and pre-warm window under a probabilistic SLA constraint. It further uses a hybrid hit-ratio estimator that combines empirical reuse profiling with model-based extrapolation and calibrates itself to workload skewness. We implement Arrietty in an extended trace-driven cache simulation framework and evaluate it using real-world storage traces and controlled synthetic workloads. The results show that Arrietty achieves a better trade-off among memory cost, backend I/O, and SLA violation risk than coarse-grained proactive baselines, especially under bursty and uncertain access patterns.

Anti-money laundering (AML) in blockchain ecosystems is an inherently adversarial task: detection models must keep pace with laundering typologies that evolve faster than supervised retraining can absorb. Graph neural networks and GNN-language-model hybrids have advanced AML detection on transaction graphs, yet they typically treat AML as a static classification problem and respond to new evasion patterns by expanding labeled corpora rather than by adapting to the adversary. We argue this gap calls for a co-evolutionary view: detector and adversary should sharpen each other within a closed loop. We instantiate this view as Fin-ASAEC, in which an Adversarial Financial Data Generator pairs a conditional Wasserstein GAN with a PPO-based strategic optimizer to synthesize diverse laundering typologies, and a Multimodal Financial Anomaly Detector turns its scores into both PPO rewards and an adversarial regularization term that hardens the target model. On a 157K transaction benchmark spanning five blockchain networks (Ethereum, Bitcoin, BSC, Solana, Polygon), Fin-ASAEC raises adversarial AML detection from 80.1% (fine-tuned baseline) to 95.2%, with consistent margins of over 11 points on unseen attack families (PGD-Graph, Nettack). The results indicate that co-evolutionary hardening offers a workable alternative to AML defenses built around data expansion.

The proliferation of fake news poses a persistent challenge to the reliability of Web information, especially when misinformation involves recent events, incomplete evidence, or deliberately misleading narratives. Although large language models have shown promising capabilities in textual understanding and reasoning, they still suffer from knowledge obsolescence, insufficient evidence grounding, and vulnerability to superficial or biased reasoning. To address these limitations, we propose RAAR, a retrieval-augmented adversarial reasoning framework for zero-shot fake news detection. RAAR decomposes the verification process into three stages. First, an evidence retrieval and synthesis module formulates verification-oriented queries, retrieves relevant external evidence, and condenses it into structured support for downstream reasoning. Second, a multi-faceted claim analysis module examines the news content and retrieved evidence from complementary perspectives, including factual consistency, source reliability, semantic coherence, and contextual plausibility. Third, an adversarial reasoning and adjudication module stress-tests the initial analysis through opposing argumentative roles and produces a final veracity judgment. Experiments on benchmark fake news detection datasets, including Weibo and GossipCop, show that RAAR improves zero-shot detection performance over strong prompting and retrieval-based baselines. Ablation studies further demonstrate the contribution of retrieval synthesis, multi-perspective analysis, and adversarial adjudication. These results suggest that combining Web-based evidence retrieval with structured adversarial reasoning is an effective direction for reliable and explainable fake news detection.

Abstract. Airborne LiDAR bathymetry (ALB) provides an efficient technique for rapidly acquiring nearshore bathymetry and underwater topographic information. However, accurate extraction of water-surface and bottom echo signals remains challenging in complex island-reef environments, especially in extremely shallow areas where surface and bottom echoes are strongly overlapped and in deeper areas where bottom echoes are weak and easily affected by water-column scattering and system noise. To address these problems, this study proposes a multi-channel time-frequency attention fusion method, named MCTFA, for full-waveform airborne LiDAR bathymetry. The proposed method uses the full-waveform data from Channels 2, 3, and 4 of the Mapper5000 system as model inputs. First, empirical mode decomposition is applied to each channel to extract multiscale waveform components. Then, the Hilbert-Huang transform is used to construct time-frequency energy maps, which enhance the separability of overlapping shallow-water echoes and weak deep-water bottom echoes. Finally, a multi-channel attention fusion network is designed to adaptively integrate complementary information from different receiving channels and predict the temporal positions of water-surface and bottom signals. Experiments conducted over Ganquan Island show that the proposed method achieves an RMSE of 0.50 m, an R² of 0.99, and a fitted slope of 0.95 against SHOALS-3000 reference data. Compared with the comparison method, the RMSE is reduced by 0.19 m, demonstrating improved bathymetric accuracy and robustness in nearshore reef environments.

Remote sensing spatiotemporal fusion integrates complementary spatial and temporal information from heterogeneous sensors to reconstruct fine-resolution imagery at a target date, supporting continuous and detailed monitoring of abrupt disasters such as floods. However, under substantial land-surface change, the reliability of historical fine structures varies spatially, making it difficult for existing methods to preserve valid details in stable regions while updating obsolete structures in abrupt-change regions. To address this problem, we propose CADD-STF, a change-aware dual-diffusion framework in which spatial and temporal diffusion streams independently generate complete target-date candidates. The spatial stream reconstructs current spatial structure from the target-date coarse observation while suppressing unreliable historical details in abrupt-change regions, whereas the temporal stream transfers coarse-observed temporal variation on the basis of the historical fine image to preserve reliable details in stable and gradually changed regions. A change-state estimator identifies stable, gradual, and abrupt regions, and an adaptive fusion module combines the two candidates after generation using target-date evidence and multiscale blending. Experiments on two regions from the 2024 Rio Grande do Sul flood event show that the proposed method consistently outperforms traditional and learning-based spatiotemporal fusion approaches, more accurately reconstructing flood extent and boundaries while preserving fine spatial structures in unaffected areas. Ablation studies further demonstrate the complementarity of the two diffusion streams and the effectiveness of change-state guidance and adaptive fusion.

Rapid flood delineation supports emergency response, but cloud-only processing incurs communication costs, while compact edge models often lose shoreline details. Metric-Aware Boundary Gradient Distillation (MBGD) combines global probability transfer with teacher-guided gradients within ground-distance bands derived from GeoTIFF metadata. Area-Consistent Boundary Refinement (ACBR) applies a boundary-gated residual with area and perimeter regularization in map units. An uncertainty-aware WebGIS routes difficult tiles to a cloud ensemble while preserving provenance information and polygon review records. On the BlessemFlood21 test split, three independent runs of the full model achieve average scores of 0.8977 ± 0.0067 IoU, 0.9461 ± 0.0037 F1, and 0.7874 ± 0.0093 BF1@1 m. Compared with the matched plain MobileNetV3 student, these results correspond to improvements of 2.2%, 1.2%, and 12.8%, respectively. Relative to the strongest listed external lightweight baseline, the corresponding gains are 3.2%, 1.7%, and 2.3%. Routing 10% of the 463 test tiles increases IoU from 0.9049 to 0.9202 and reduces raw data upload by 90.1% compared with full-cloud processing.

Reliable rice field segmentation is a prerequisite for disaster-oriented agricultural monitoring and crop damage assessment. Although UAV imagery and ground spectral measurements provide valuable complementary information for hyperspectral rice field segmentation, their temporal inconsistencies and heterogeneous spatial coverage make direct multi-source fusion difficult. To address this limitation, we propose MSPNet, a multi-source privileged learning framework for hyperspectral rice field segmentation. MSPNet takes only a 32-band hyperspectral image (HSI) during inference while exploiting UAV imagery and ground spectra as training-time privileged information. MSPNet introduces a Prototype-Adaptive Spectral Prior (PASP) module, which calibrates ground spectral prototypes with image-level class spectral centroids to generate adaptive spectral similarity maps that are concatenated with the hyperspectral input before encoding. For UAV guidance, MSPNet further introduces a Coverage-Aware Boundary Guidance (CABG) module, where high-resolution UAV-derived boundaries are applied only within valid UAV-covered regions to enhance field-boundary learning without requiring UAV imagery during inference. Experiments on the Anxiang Hyperspectral Dataset demonstrate that MSPNet achieves an OA of 79.92%, an mIoU of 41.92%, and an mF1 of 53.11%, outperforming several representative segmentation methods. Ablation studies further demonstrate the complementary benefits of PASP and CABG.

Emergencies at sea are shaped by sea conditions, target status, and communications. Ongoing changes cause updates to mission creation, resource allocation, operational strategy, and field execution. Existing frameworks often struggle to align corporate strategy with at-sea operations, resulting in discrepancies in plan iterations and responses. The following paper details a hierarchical, collaborative system designed for planning and control in maritime emergency situations. This system comprises an Action Planning and Management System (APMS) and a Maritime Emergency Mission Planning and Management System (MEMPMS). APMS includes a shore-based subsystem for managing data acquisition missions and a ship-borne subsystem for executing them. It handles emergency missions, improves operational plans, aligns ship and shore data, organizes equipment, oversees execution, and sends back gathered data. The MEMPMS comprises a General-Purpose Mission Processing Subsystem, a Maritime Emergency Control Mission Management Subsystem, and a Maritime Emergency Response Mission Management Subsystem. It handles urgent needs, oversees resources and limitations, starts planning computations, examines plans, and assists with control in specific situations. Assigning duties and connections allows for phased integration and testing.

AI-driven Web-based flood emergency information systems need to integrate heterogeneous knowledge from emergency plans, policy documents, and historical disaster records and transform it into normative, scenario-specific, and evidence-traceable actions. However, existing retrieval-augmented generation methods often treat heterogeneous evidence as homogeneous context, without distinguishing the roles of policy clauses and historical experience or ensuring that retrieved evidence is fully realized as actions. This paper proposes PlanRealizeRAG, a policy-guided evidence-to-action framework for flood emergency decision support. It separately retrieves policy clauses and historical experience, uses policy action semantics and applicability conditions to guide historical-event ranking, and constructs a policy-grounded structured action plan. Relevant historical experience is then incorporated to supplement risk objects, failure-prevention information, and execution details, while coverage auditing and slot preservation reduce action omissions and retain evidence provenance. Experiments on 126 explicit and implicit queries show that PlanRealizeRAG achieved policy-clause and historical-event Hit@1 scores of 0.849 and 0.444, respectively. It also obtained an ActR of 0.758, an ActRec of 0.855, and a context-valid citation rate of 1.000, demonstrating improved heterogeneous-evidence retrieval, action-information coverage, and evidence provenance preservation.

China's edible fungus industry has emerged as a significant contributor to ru-ral revitalization and farmers' income growth. As a major production hub, Hu-bei faces challenges such as information fragmentation, data silos, insufficient spatial visualization, and scattered technical services during industrial expan-sion, which hinder effective production management and decision-making. Ex-isting agricultural information platforms are mostly designed for general agri-cultural scenarios and lack sufficient capabilities in multi-source data integra-tion, spatial analysis, and decision support tailored to the mushroom industry. To address these issues, this paper designs and implements a digital manage-ment and analytical platform for Hubei's mushroom industry. The platform in-tegrates heterogeneous data sources—such as policies and regulations, germplasm resources, production and marketing time-series, production area spatial data, and cultivation techniques—with dynamic updates enabled by au-tomated web crawling, and achieves multi-dimensional presentation through GIS-based spatial analysis and statistical visualization. The backend adopts a lightweight microservice architecture, the data layer employs a distributed document database for unified and efficient processing of heterogeneous data, and the frontend utilizes component-based development combined with GIS maps and interactive charts to present industrial distribution, production trends, and variety composition from provincial to municipal scales. The sys-tem is complemented by cultivation guidance and backend administration modules, covering the entire information service workflow. Testing and prac-tical deployment demonstrate that the system offers strong stability and practi-cality in data integration, visualization, and technical service delivery, provid-ing solid support for information-based governance, data-driven decision-making, and intelligent services for Hubei's mushroom industry.

Global biological invasions have intensified, with over 140 invasive plant spe-cies documented in Hubei Province, inflicting substantial agricultural and eco-logical losses. A preliminary province-wide questionnaire revealed that nearly 80% of respondents lacked basic knowledge of invasive alien plants, with only 5% indicating that they were "very familiar" with the issue. Current prevention and control efforts in Hubei rely primarily on manual field surveys, with lim-ited public participation leading to high costs and low response efficiency. To address these problems, this study develops a digital platform for the preven-tion and control of invasive alien plants in Hubei Province. The system inte-grates public education, AI-assisted species identification, visualization of in-vasion dynamics, collaborative public reporting, and end-to-end regulatory tracking. The platform offers three core advantages. Regional specificity is supported by a dedicated database covering the local distribution and impact characteristics of over 140 invasive plant species in Hubei. Workflow coordination enables one‑step photographing, AI‑based identification, and location reporting, while simultaneously delivering science communication, thus creating a seamless tripartite linkage among government, research, and public stakeholders. Real-time decision support allows administrators to promptly assess distributions and devise control measures, forming a complete closed loop: detection and identification, reporting and verification, treatment tracking, and decision optimization. The system has been validated as stable and practical in multi-source data integration, dynamic visualization, and multi-stakeholder collaborative governance. The platform directly supports information-based governance, scientific control, and public participation in Hubei, effectively enhancing regional management efficacy with strong potential for broader applicability and regional demonstration.

Rainfall-induced landslides can degrade road performance, trigger route reconstruction, and disconnect settlements from administrative centers. Existing GIS-based and network analysis methods can quantify these effects, but their outputs are usually distributed across spatial layers, network files, route sequences, and accessibility tables, making it difficult to trace accessibility consequences back to specific road segments and landslide objects. This paper proposes a scenario-aware path-level knowledge graph for integrating landslide–road spatial relations, road disturbance scenarios, pre- and post-disturbance routes, and settlement accessibility outcomes. The proposed model represents settlements, towns, landslides, road segments, disturbance scenarios, accessibility results, routes, and ordered route steps as interconnected semantic entities. Three road-disturbance scenarios are designed: continuous impedance increase, local road blockage, and blockage of landslide-intersecting roads. The knowledge graph supports cross-scenario accessibility comparison and bidirectional impact tracing from landslide to settlement through road segment, route, and accessibility result. Using rainfall-induced landslides in Zixing, Hunan Province, China, as a case study, we constructed a knowledge graph comprising eight node types and thirteen relationship types, including 1,667 landslide nodes, 28 settlement nodes, four town-center nodes, and 1,069 route-step nodes. Query demonstrations show that the graph can distinguish whether settlements remain reachable, require rerouting, or become disconnected from town centers across different disturbance scenarios, compare pre- and post-disturbance route sequences, and trace accessibility consequences through disturbed road segments to the associated landslides. The proposed approach provides a structured and queryable foundation for scenario-based disaster impact assessment and evidence-informed emergency decision-making.

Electroencephalogram (EEG)-based emotion recognition and regulation in computer-assisted therapy has drawn increasing attention with the increase of emotional stress. Although the traditional tasks of EEG recognition and regu-lation have reached a certain level of classification, they are still affected by the complexity of the model, the incompleteness of feature extraction and the problem of unable to obtain the regulation effect of the current real emotions. In this article, we propose ER-ECT, the first intelligent emotional state recog-nition and regulation mechanism based on EEG to tackle these challenges. ER-ECT combines the feature extraction features of CNN and Transformer in an adaptive weighted manner to train an emotion recognition model ECT, which then fuses the EEG into the regulation cycle through the contextual-bandits to learn different regulation habits of different individuals, thus real-izing a relatively personalized regulation model. The SEED data and emotion regulation experiments have evaluated that ER-ECT has good recognition accuracy and regulation effectiveness, which is feasible and extensible in computer-aided applications.

Industrial Control Systems (ICS) in critical infrastructures are vulnerable to sophisticated cyberattacks. Pure data-driven anomaly detection approaches often overlook the physical constraints and temporal dynamics inherent in industrial processes, which may degrade the detection performance. This paper proposes a physics-informed and data-driven framework that combines Graph Attention Networks (GAT) with deep temporal analysis for robust ICS anomaly detection. The key innovation lies in a composite loss function that jointly optimizes three complementary objectives by uncertainty-based weighting: (1) point-wise reconstruction via Mean Squared Error (MSE), (2) temporal consistency through MSE-based penalties on velocity (first-order) and acceleration (second-order) derivatives to enforce smooth temporal evolution, and (3) dependency-aware regularization that penalizes inconsistent prediction errors across topologically connected sensors in the physical-informed graph. It is proven to be helpful for the detection performance.

The increasing deployment of heterogeneous sensing infrastructures in critical infrastructure and cultural heritage monitoring has shifted the primary engineering challenge from data acquisition to developing interoperable, intelligent, and trustworthy decision-support services. Existing monitoring platforms typically integrate sensing, analytics, and visualisation within application-specific architectures, limiting their reusability, scalability, and interoperability across different engineering domains. This paper presents AGENTSAFE, a service-oriented framework that transforms multi-source monitoring data into trustworthy engineering decisions through a modular orchestration of interoperable services. Rather than treating sensing devices as the principal contribution, the proposed architecture decomposes the entire monitoring workflow into independent service layers: Sensor Services, Data Acquisition Services, Data Preprocessing Services, Decision Support Services, Risk Assessment Services, and Notification Services. Each service exposes standardized interfaces, enabling independent deployment, composition, replacement, and reuse within heterogeneous cyber-physical environments. The AI service layer (Decision Support and Risk Assessment) includes an improved approach that reduces the risk of overfitting, making the model simpler and more robust. The AGENTSAFE architecture is validated in a cultural heritage monitoring context, where distributed environmental, hydrological, geotechnical, and structural measurements are orchestrated as reusable services to support continuous monitoring and predictive risk assessment. Experimental results demonstrate that the service-oriented decomposition improves modularity, interoperability, scalability, and maintainability while enabling seamless integration of AI-based analytical services without modifying the underlying sensing infrastructure.

Global food supply chains require timely and systematic monitoring of food safety risks. The Rapid Alert System for Food and Feed (RASFF) provides valuable information on food products, hazards, countries of origin, and notification types; however, heterogeneous terminology limits the consistent analysis of historical food safety data. This study proposes an automated large language model (LLM)-based Retrieval-Augmented Generation (RAG) framework (RAG-LLM) to standardize product and hazard labels in RASFF notifications. The framework integrates domain-specific fine-tuned LLM predictions with top-$k$ standardized product and hazard labels retrieved from a vector database, enabling dynamic prompting based on relevant standardized labels and historical examples. Human validation, iterative vocabulary updates, and recursive fine-tuning are incorporated to continuously improve the standardization process. Experimental evaluation demonstrates that integrating semantic retrieval from domain specific dictionaries, domain-specific fine-tuning, and a stronger reasoning model achieves more accurate and consistent product and hazard standardization than the standard LLM-based methods. The proposed framework therefore provides a foundation for automated food safety monitoring for food recall alerts and can be extended to other similar food safety data resources.

MethodVerse is an AI-driven platform designed to make research methodologies easier to use and reuse across different datasets and research contexts. Many research methods are often stored as separate scripts, require specific configurations, and demand technical expertise, which can limit their accessibility. MethodVerse addresses this challenge by turning research methods into modular services that can be connected through a drag-and-drop visual workflow interface. Researchers can select data sources, configure methodology nodes, run workflows, and generate results without manually writing or connecting code. By lowering the technical barriers to applying AI-driven analytical methods, MethodVerse can support researchers studying socially relevant Web phenomena such as online behavior, recommendation systems, misinformation, and public discourse. The system supports workflows such as YouTube recommendation analysis, where data moves from MongoDB inputs through recommendation graph generation, metadata extraction, analysis, and AI-driven insight generation. Using large language models, MethodVerse also produces readable summaries, evidence-based explanations, reports, and presentation-ready findings. Overall, MethodVerse provides a unified environment for operationalizing research methods and is designed to lower technical barriers by helping researchers move from raw data to interpretable outputs through reusable visual workflows.

Keratoconus is characterized by the thinning and bulging of the cornea which leads to impair vision and distortion. Research highlighted the potential correlation between eye rubbing and the development of Keratoconus. The reviewed literature found that existing AI-based tools that detect pathological and post-surgery eye rubbing are generally either inaccurate or platform specific. To cope with this problem, we propose a custom Transformer-based model with attention mechanisms and time-series data for the classification of eye rubbing behavior in individuals. A Samsung Galaxy Watch 4 smartwatch was used for data collection, utilizing its onboard sensors. Having achieved excellent results from the training and the implemented application on the smartwatch, the detection of eye rubbing was successfully accomplished obtaining a test accuracy of 98.14%.

Rapid assessment of critical infrastructure is essential after disasters, particularly in mountainous and remote areas where ground access may be restricted. UAVs can follow prescribed inspection routes to collect situational information, but limited onboard energy and time-consuming charging detours may delay mission completion. This paper formulates a fixed-route UAV charging scheduling problem that determines when to initiate charging, which ground station to visit, and how much electricity to charge. The objective is to minimize mission completion time while maintaining battery safety. The model accounts for phase-dependent energy consumption during ascent, cruise, and descent, together with the practical charging feature that the charging speed is nonconstant during the charging process. We develop a triple-tuple encoded genetic algorithm (TTGA) that combines coupled decision encoding, tuple-level genetic operations, and feasibility-aware decoding and repair. Experiments on nine synthetic terrain instances with normalized parameters compare TTGA with the greedy charging strategy. TTGA reduces the mean mission completion time by 50.7\%--64.9\% relative to greedy baseline under the reported settings. These results indicate the potential of intelligent charging scheduling to shorten the time required to complete prescribed UAV inspection routes and thereby support more timely post-disaster situational assessment.

The rapid democratization of Large Language Models (LLMs) has amplified a dual-use dilemma: the same capabilities that let these models advance legitimate science also let them generate unsafe content in domains such as biosecurity and chemistry. At the same time, the demand for privacy-preserving training has driven the adoption of Federated Learning (FL), in which a central server aggregates client updates but never sees the raw local data. This privacy guarantee disables the dominant safety paradigm: server-side inspection and filtering of training data becomes impossible, so a client fine-tuning on hazardous data can silently raise the hazardous capability of the global model. Our key observation is that, although the server cannot select data, it can still act on a standardized risk signal that each client computes locally and shares without ever exposing and dropping raw samples. The goal of this work is therefore to relocate safety enforcement from the data space into the aggregation space, using this risk signal to properly weight the incoming updates. Building on a Parameter-Efficient Fine-Tuning (PEFT) approach, we evaluate a spectrum of aggregation strategies, including a novel Federated Risk-Aware Projection (RAP-FL). Across three client data distributions, we quantify the trade-off between suppressing hazardous knowledge (WMDP) and preserving general-domain utility (MMLU), and find that RAP-FL gives the most favorable safety-utility balance among the strategies we tested.

Artificial Intelligence (AI) in healthcare depends on sensitive patient data, creating substantial privacy risks. Differential Privacy (DP) provides formal protection by limiting what model outputs reveal about any individual, but its practical use in healthcare remains unclear. We systematically reviewed 25 empirical studies published between 2022 and 2025 that applied DP to healthcare data with machine-learning (ML) models. We examined where DP was integrated, how privacy budgets were selected and reported, and how privacy-utility trade-offs were evaluated. DP was used across clinical prediction, medical imaging, wearable-data analysis, synthetic-data generation, and federated learning. However, implementation and evaluation were inconsistent. Privacy parameters were rarely linked to clinically meaningful risks or threat models, mechanism selection was often driven by model compatibility, and evaluations emphasized aggregate predictive performance more than privacy attacks, subgroup fairness, or deployment readiness. These findings show that DP is technically feasible in healthcare AI but is not yet supported by consistent, clinically grounded practices. Future work should connect formal guarantees to realistic disclosure risks, justify privacy budgets and mechanisms, assess subgroup effects, and validate DP-enabled systems in operational clinical environments.

World models take multimodal inputs like text, photos, and diagrams to generate dynamic scenes in accordance with the laws of physics, thus opening a compelling application: fusing multimodal legal evidence to re-create a crime scene and re-enact how an offence could have been committed. However, feeding the raw, unorganized evidence into a world model fails in forensic use: it silently drops evidence, glosses over contradictory testimony, and produces motion that violates the evidentiary record. This paper presents VeriScene, an agent that orchestrates the world model: it reconstructs crime scenes from forensic photographs and witness statements of varying reliability, keeping every claim traceable to evidence and every motion physically plausible. VeriScene iteratively fuses the evidence into a cited narrative under an auditing loop, verifies the hypothesized dynamics via probe rollouts in the world model with corrective constraint injection, and renders the offence as a re-enactment video from a fused keyframe. On a benchmark of 25 crime scenarios across 7 physically-driven case types (139 forensic-style photographs and 65 statements with planted unreliability), VeriScene attains 0.9014 evidence coverage and 0.7217 factual consistency (0-1 scale) on the 20 test scenes, outperforming an end-to-end multimodal-LLM baseline by 20.35% in factual consistency and 34.88% in temporal coherence, while generalizing across four LLM orchestration backends at $1.82 per scene.

Bid optimization in online advertising relies on black-box machine learning models that learn bidding decisions from historical data. However, these approaches sometimes fail to replicate human experts’ adaptive, experience-driven, and globally coherent decisions. Specifically, they generalize poorly in data-sparse cases because of missing structured knowledge, make short-sighted decisions that ignore long-term interdependencies, and struggle to adapt in out-of-distribution scenarios where human experts succeed. To address this, we propose KBD (Knowledge-informed Bidding with Dual-process control), a novel method for bid optimization. KBD embeds human expertise as inductive biases through the Informed Machine Learning paradigm, uses Decision Transformer (DT) to globally optimize multi-step bidding sequences, and implements dual-process control by combining a fast rule-based PID (System 1) with DT (System 2). Extensive experiments highlight KBD’s advantage over existing methods and underscore the benefit of grounding bid optimization in human expertise and dual-process control.

Public pre-trained model hubs have become a primary supply channel for production machine learning, and the openness that powers their adoption has also turned them into a fast-growing attack surface. Pickle-scanning bypasses (CVE-2025-10155), namespace reuse attacks against Vertex AI and Azure AI Foundry, and more than one hundred malicious models on the Hugging Face Hub in 2024--2025 have moved AI supply-chain attacks from theoretical to operational. We define the defense problem as anomaly detection on a heterogeneous, lineage-aware, and temporally evolving graph that ties models, datasets, files, users, and organizations, and we present HALO-MAD, Heterogeneous Attack-aware Lineage-Oriented Model Anomaly Detection, a single unified detector that targets four high-impact attack archetypes (pickle payload, namespace reuse, typosquatting, and dormant-account hijack) in one publish-time inference pass. HALO-MAD pairs neighborhood reconstruction with two application-tuned specialist channels (typosquat tail-product and lineage-conditioned propagation), then fuses all channels into a single calibrated score. An Attack-Pattern-Con\-di\-tioned Augmentation library synthesizes positives consistent with public threat intelligence for stress testing and supervised retraining, an Adversarial Card Robustness step discounts attacker-controllable model-card text, and calibrated triage cards give security analysts decomposable evidence. On the synthetic Hub graph HALO-MAD outperforms six graph anomaly baselines by 7.13 and 11.80 percentage points on AUPRC and P@n; on the real Hugging Face Hub graph it outperforms the strongest overall baseline (Isolation Forest) by 75.17 and 88.65 percentage points on AUPRC and P@n, with the ranking preserved across three subsamples.

Question Answering (QA) systems often struggle with complex questions that demand synthesis of knowledge across multiple domains. While naive Retrieval-Augmented Generation (RAG) models excel at basic information retrieval, they fall short when comprehensive and in-depth explanatory answers are required. To address this gap, we introduce the task of explanatory answer generation, which emphasizes both breadth of information and logical coherence. Inspired by systematic thinking theory, we propose SynthRAG—an adaptive framework that enhances QA performance through three key mechanisms: dynamic outline generation for structured content, systematic information synthesis for thorough coverage, and customized answer generation tailored to user queries. This approach ensures logically coherent and well-integrated responses that are both insightful and methodically organized. Empirical results demonstrate SynthRAG’s superiority in handling complex questions, significantly outperforming existing RAG-based methods in answer quality and depth. In an online deployment on Zhihu, SynthRAG achieved an average of 5.73 upvotes per answer, surpassing 79.8% of human contributors, underscoring its practical impact and relevance.

Enterprises are beginning to use LLM agents to automate business intelligence (BI), but deployed platforms such as Apache Superset still rely on browser-centric workflows for dataset selection, chart configuration, dashboard assembly, and monitoring. The difficulty is not just query generation. Agents must create governed, persistent artifacts through APIs partly defined by frontend code. In Superset, accepted calls may still yield empty charts, misplaced dashboards, or broken scheduled views. We report an internal deployment that makes Superset agent-operable without modifying its codebase. The retrofit combines an executable skill for hidden artifact logic, a verification-and-repair loop over SQL, chart payloads, layout, and rendered outputs, and an HTTPS interceptor that preserves per-user audit under enterprise SSO. The agent discovers datasets, authors charts and dashboards, schedules monitoring, and sends summaries through enterprise messaging. We evaluate the design with a 90-task NL2BI benchmark and an internal deployment audit. The skill-guided agent reaches 85.4% workflow success with no observed field hallucinations, outperforming direct and schema-grounded LLM baselines. In deployment, it serves roughly 80 analysts and produces 433 charts, 58 dashboards, 16 monitored dashboards, and 99.9% IM delivery. The results support this retrofit as a practical pattern for multi-step BI workflows on frontend-coupled BI systems.

Climate accountability work, carried out by investigative journalists, NGOs, and watchdog organisations, depends on knowing who lobbies against climate action and how these actors are connected. This evidence is documented in depth by platforms such as DeSmog and LobbyMap, but remains fragmented, heterogeneous, and hard to analyse at scale. We present LACK (Lobbying Agents, Climate, and Knowledge), to our knowledge the first openly published knowledge graph of climate lobbying networks: ~28k entities and ~47k asserted relations (~189k after inference), built from over 2k web pages. By making these relationships queryable, LACK is designed to support the tasks practitioners perform: tracing funding flows, tracking how actors' positions shift over time, and mapping influence networks across organisations. We demonstrate LACK through case studies, report how domain experts assess its value, and discuss the opportunities and challenges of building such resources automatically in this domain.

Large language model (LLM) agents increasingly depend on external memory modules to support long-term, multi-session interactions. However, existing memory retrieval in A-Mem mainly ranks stored memories by semantic similarity, leaving timestamps underused and making recent but time-critical information difficult to prioritize. This limitation is particularly problematic in practical conversational agents, where user preferences, task states, and factual records may change over time. This paper presents A-Mem$^*$, a lightweight temporal retrieval extension to A-Mem that improves memory freshness without model fine-tuning. A-Mem$^*$ injects normalized timestamps into memory embedding inputs and further applies a query-aware temporal prompt with exponential decay-based re-ranking. The method is simple to implement, compatible with existing vector indexes, and configurable through the decay factor $T$ and time weight $\alpha$. Experiments on the LoCoMo benchmark show that A-Mem$^*$ improves F1 by 15.3\%, 9.3\%, and 6.9\% over A-Mem on MultiHop, SingleHop, and Adversarial tasks, respectively, while the retrieval-side variant A-Mem$^t$ further improves Temporal-task performance. These results demonstrate that temporal signals can provide practical gains for deployable long-term memory systems.

We present CoRIE (Context-Restricted Information Extraction), an industry-deployed system for extracting structured enterprise knowledge from company websites using a context-restricted large language model (LLM) pipeline. CoRIE combines depth-limited webcrawling, page-type classification via a fine-tuned transformer, and targeted LLM prompting over relevance-filtered content to improve robustness, cost efficiency, and scalability in open-web information extraction. Deployed within an enterprise data-enrichment workflow, the system processes up to 500,000 company domains per day across diverse industries and regions while handling heterogeneous, frequently changing website structures. On a manually annotated gold benchmark of 2,500 company domains, CoRIE attains 97% exact-match accuracy on structured attributes, compared with ≈66% for the strongest unrestricted-LLM baseline (a relative improvement of ≈47%), together with consistent gains on multi-entity (+51.5% F1), open-vocabulary semantic (+8.2% cosine), and free-text (+61.8% BERTScore) attributes. A component ablation shows that page-type-conditioned context restriction, rather than model scale, is the dominant factor governing extraction reliability, while reducing end-to-end cost by 66–89%. We additionally report the attribute taxonomy and its construction, the gold-standard annotation protocol, an error taxonomy with measured frequencies, and downstream enterprise-impact statistics. These results demonstrate that context-restricted LLM pipelines provide a practical foundation for scalable, cost-effective enterprise web intelligence.

Rapid response is critical for mobile search systems, where even small delays can reduce user engagement and conversion. Existing latency reduction methods, including server-side optimization and model compression, have delivered limited additional gains in large-scale production environments. This paper presents TRIP (Transfer-based Retrieval Intent Prediction), a lightweight on-device framework for predicting search intent from user typing behavior. By modeling the dynamics of text input in real time, TRIP identifies moments when the user is likely to pause and initiates retrieval before the query is formally submitted. This design leverages natural typing interruptions to overlap back-end retrieval with the input process and reduce perceived latency. Experiments on a large-scale e-commerce mobile platform show that TRIP achieves a precision of 0.7953, a recall of 0.8218, and an accuracy of 0.7758. In production, the system reduces average latency by 635 milliseconds and increases Gross Merchandise Value by 0.34 percent. The framework also supports sub-millisecond inference and imposes only minimal overhead on-device resources. These results suggest that moving intent prediction to the edge is an effective approach for improving responsiveness in real-time mobile search applications.

The operational memory of maintenance technicians constitutes a strategic asset that is rarely formalized. In industrial environments, this knowledge, accumulated through years of field experience, is primarily transmitted orally and remains highly vulnerable to staff turnover. This thesis aims to create a structured memory of the heterogeneous data associated with industrial maintenance interventions, e.g., involving causal, temporal, and procedural information. In particular, we want to design and evaluate methods based on LLM agents that manage and access an integrated memory of maintenance interventions in the form of a knowledge graph. To achieve this, we address several research challenges: assess the memory forms supported by current memory systems w.r.t. requirements of industrial maintenance; design a memory model adapted to industrial entities and sequences; automatically extract, update and use this memory in the agents; and evaluate the methods in a real-world context. In our preliminary work, we have surveyed the various forms of memory in order to compare features of existing memory systems with the requirements of industrial maintenance, which combines abstract concepts with concrete interventions. This led us to propose a taxonomy of memory forms unifying definitions coming from the fields of cognitive science, knowledge management, and artificial intelligence.

Unified Extensible Firmware Interface (UEFI) Secure Boot restricts pre-OS execution to cryptographically signed components. Recent Secure Boot disclosures illustrate that cryptographic signature verification alone does not guarantee the security of the boot chain when trusted binaries contain exploitable logic or when trust material is incorrectly provisioned. We review six recent Secure Boot-related cases and organize them according to the layer at which the security guarantee fails: trusted-binary behavior, trust-material provisioning, and firmware implementation flaws. Our review did not identify a targeted assurance step at the transition from a compiled EFI artifact to a production signing or release decision that evaluates these trust-boundary properties explicitly. The proposed design examines two classes of security indicators: compiled-binary characteristics associated with manual image-loading behavior and known-bad or test Platform Key material present in build or provisioning artifacts. The contribution of this paper is analytical and methodological rather than empirical. The paper does not present a validated detection system. Instead, it identifies key risk indicators, analyzes gaps in current approaches and proposes an artifact design together with a reproducible evaluation framework. The indicators are intended to identify artifacts that merit further security review rather than to provide a binary vulnerable/safe classification.

Multi-omics data are fragmented across public repositories, described by sparse and inconsistent metadata, and produced by workflows whose outputs do not fit the table-based formats used to share them. This fragmentation makes the data hard to find, reuse, and query jointly, which blocks molecular exploration. Knowledge graphs and artificial intelligence (AI) offer a promising approach to integrate such disparate, multimodal omics data: large language models (LLMs) are emerging as a new way to extract the biological context of an experiment, while knowledge graphs contextualize and integrate the resulting data. Relying solely on LLMs is problematic, as they can hallucinate, and a knowledge graph can be corrupted. The purpose of this thesis is to investigate ways to build a faithful, provenance-aware multi-omics knowledge graph from several public repositories. It asks three questions: can human oversight and an LLM collaborate to build such a graph; can disparate, heterogeneous, and large biological data be integrated into a single, maintainable knowledge graph; and can an LLM be grounded in a knowledge graph to answer biological questions faithfully and completely. Current AI-for-science systems mostly synthesise the literature and rarely connect to the underlying measurements. This work will deliver a framework called MultiOmicsKG, in which a constrained linguistic agent bridges that gap, so that biological researchers can obtain actionable insights from multi-omics data for AI-assisted molecular exploration.

Knowledge-graph-grounded large language models (LLMs) are usually evaluated with the graph treated as reliable. We instead study what happens after retrieval when the evidence itself is wrong. Starting from a validated 300-question MetaQA pool, we construct a matched evaluation core of 28 multi-hop questions (14 two-hop and 14 three-hop) that is feasible for four controlled corruptions: entity substitution, relation substitution, contradictory insertion, and intermediate-node rerouting. Each condition is evaluated with natural entity names and with deterministic anonymized identifiers while the question, prompt, model, and decoding setup remain fixed. In the Qwen3-4B evaluation, entity substitution causes the largest drop: exact-match accuracy falls by 50.0 percentage points with natural labels and 46.4 points after anonymization, reaching 0% in both cases. Relation substitution and contradiction also cause large losses, whereas rerouting is less damaging. We find no clear increase in corruption sensitivity after anonymization: all Parametric Masking Gap (PMG) estimates are non-positive and every 95% paired-bootstrap interval includes zero. Under contradiction, the model returns both the gold and injected alternatives in 60.7% of natural-label cases and 75.0% of anonymized cases. These results show that clean-answer accuracy alone is insufficient to establish faithful graph use.

Timely identification of the acute phase of myocardial infarction (MI) from the electrocardiogram (ECG) is critical for emergency clinical decision-making, since outcomes are strongly time-dependent. Existing deep learning studies on the PTB-XL dataset largely target MI detection at the superclass level or localization by anatomical subclass (anterior, inferior, lateral, posterior), while overlooking a dataset-provided but underused pair of fields, infarction_stadium1 and infarction_stadium2, which encode the clinical stage of the infarction (acute, subacute, or old/stabilized) as extracted by the original annotators from the cardiologist reports. This paper proposes a hybrid convolutional–Transformer architecture that learns to jointly detect acute/recent MI and localize its anatomical region directly from raw 12-lead ECG signals, without hand-crafted morphological rules. Acuity ground-truth labels are constructed from the infarction_stadium fields, cross-validated against free-text report keywords, and used solely for supervision and evaluation — never as model input — so that the network must learn to recognize the relevant ST-segment and Q-wave morphology on its own. We describe the dataset preparation pipeline, the proposed multi-task architecture, the training protocol, and an interpretability strategy based on Grad-CAM and attention-weight visualization intended to verify that the model’s learned representations align with established electrocardiographic acute-MI signatures. This work is presented as ongoing research at an early-to-intermediate stage, with the described pipeline implemented and preliminary experiments underway.

Cross-library search for formal mathematics is commonly evaluated against a single accepted target, although independently engi neered libraries may split the same source content across di erent mod ule boundaries. We test whether this organizational mismatch requires graded, set-valued retrieval and whether it can be exploited without changing an existing top result. Our open-corpus study searches from 106 MathComp/Rocq modules into 7,661 mathlib/Lean modules. In a 533-pair diagnostic pool, historical pre x labels accept 78 weak or unre lated pairs and reject 41 substantial or direct pairs. Under a blinded pro tocol whose selector and evaluator use GPT-5.6 Sol in separate sessions, 52 of the 62 curated queries have structured non-single organization and 50 have multiple substantial ideal targets. We introduce primary preserving complement expansion: a xed lexical dense reciprocal-rank fusion supplies rank one, and additions are admitted only for important aspects that it leaves uncovered. The rule is developed on 20 queries, frozen, and evaluated on 24 disjoint modules outside the historical bench mark. It leaves Direct@1 unchanged at 0.125 (n = 24), raises nDCG@10 from 0.636 to 0.668 (n = 22), and raises substantial-ideal coverage@10 from 0.783 to 0.938 (n = 18); 10 of 14 additions are substantial or direct. External protocol calibration recovers 149 of 150 masked, score-marked and conservatively axiom-screened generated RocqLean counterparts at rank one. These pool-relative, model-judged results support compact complementary retrieval as an alternative to forced single-module links.

We present ContextBox, a contextual exploration interface for Wikipedia that uses usage contexts derived from Q&A data. ContextBox presents these contexts alongside Wikipedia articles, links them to relevant passages, and supports access to related information through characteristic terms. This demonstration highlights how Q&A-derived contexts can enrich Wikipedia reading and support contextual exploration.

Variable-length path queries are a fundamental query type in graph database systems. Many real-world graphs incorporate both edge labels and temporal information, such as time intervals or timestamps attached to edges. However, existing systems provide limited support for optimizing variable-length path queries that simultaneously involve relation-type constraints and edge-level temporal-span constraints. In this demo, we present SpanVLP, a system that efficiently evaluates span constrained variable-length path queries: given a source vertex s, a target vertex t, a set of allowed edge labels L, and a time interval I, SpanVLP checks for variable-length paths from s to t in which every edge uses a label in L and has a time interval contained in I. To support interac tive query processing, SpanVLP combines an online constrained traversal algorithm with a pre-built Label-Span Landmark Index (LS-Index) for suffix verification. SpanVLP also returns interpretable query evidence, including traversal-prefix edge labels and time intervals together with LS-Index-based suffix certificates. We demonstrate SpanVLP on YAGO2 and LDBC-FinBench through a Web-based interface that supports en tity search, example-based query submission, Cypher query display, and interactive result visualization.

In this demonstration, we present DecCore, an efficient dis- tributed k-core community search system for billion-scale real-world graphs.To fix the coupling issue of traditional community search algorithms, we propose a novel disentangled mechanism to split connectivity propagation and cohesion filtering into alternating execution phases. Coordinated by categorized messages and superstep scheduling, this design removes redundant transmission and pruning overhead. Built upon distributed graph processing systems, the system efficiently processes community search queries and returns reliable results. Experiments on billion-scale graphs prove that DecCore delivers low latency, high efficiency and scalability for distributed k-core community search.

Misinformation on social media threatens public trust and increases the need for effective online content governance. Although existing methods have made progress in detection and source localization, their practical deployment still faces several limitations. Many approaches assume complete network topology and sufficient propagation observations, which are often unavailable in real-world social media environments. In addition, detection and localization are usually studied independently, making it difficult to link the identification of suspicious content to subsequent intervention. To address these issues, we present MisinfoTrace, an end-to-end interactive system for misinformation governance on social media. MisinfoTrace integrates propagation modeling, key-node monitoring, veracity detection, and source localization into a unified web-based workflow. The system combines content, user-level, structural, and temporal evidence. It provides interactive views for network exploration, topology completion, key-node prioritization, propagation analysis, content verification, and candidate source inspection. Experiments on benchmark datasets and simulated propagation settings demonstrate the effectiveness of MisinfoTrace, and the system demonstration illustrates its practical utility for misinformation governance.

NoSQL systems offer multiple alternatives to represent the same dataset, making data modeling a critical design decision that affects performance, financial cost, and energy consumption. In cloud environments, evolving workloads and data characteristics can quickly render an initially optimal model suboptimal, motivating adaptive data model selection and evolution. We present DaMoOp∗, an interactive framework for cost-aware NoSQL data modeling. The framework first relies on a large-scale simulation engine that generates diverse scenarios by varying query access patterns, workloads, data distributions, and data models. This process produces millions cost observations (performance, financial & carbon footprint) which feeds a predictive model capable of estimating future costs when data characteristics or workloads change. With a baseline random forest predictor trained on simulations the TPC-C benchmark we achieve log-space R2 scores between 0.93 and 0.99. The framework also provides visual analytics to support decision-making on data model evolution. The demonstration highlights four main features: data model generation and cost simulations, details on best cost-effective candidates, alternative data models comparison, and workload-driven prediction of cost evolution with adaptive model recommendations.

Missing data imputation is critical for safeguarding data integrity, ensuring reliable data analysis, and maintaining stable performance of downstream machine learning models across diverse real-world scenarios. Existing imputation methods often overlook latent class structures embedded in unlabeled data, and they implicitly assume data homogeneity with limited robustness against inherent noise. In this paper, we present a generative adversarial framework for missing data imputation (GA4DI). This framework uses pseudo labels to guide data partitioning, thereby exploiting latent class structures, and attribute-level adversarial denoising to capture inherent feature correlations while suppressing noise. Based on this framework, we further develop an interactive demo system with visualization, hyperparameter tuning, and imputation quality assessment, and design three representative scenarios to evaluate its effectiveness under varying missing rates. Evaluation results demonstrate that GA4DI achieves lower imputation error and higher classification accuracy across most missing rates, generally outperforming mainstream neural imputation methods of the same category.

Spatial co-location pattern mining identifies subsets of spatial features that frequently co-occur within a given neighborhood area. However, under heavy-tailed feature distributions, traditional prevalence measures often suppress patterns containing rare features. We introduce RareCo, a web application for mining and visualizing these rare-feature spatial co-location patterns. The system implements a log-adaptive prevalence measure. By operating in the log-ratio domain, this method stabilizes the global dispersion estimate and recovers rare-feature associations. The web interface allows users to configure thresholds, run the clique based mining algorithm, and project the resulting co-location instances directly onto interactive maps.

Managing long-term conversational history remains a fundamental data management challenge for LLM agents. Existing memory systems primarily rely on flat vector-based semantic retrieval. This approach frequently suffers from temporal inconsistency, as it overlooks the critical chronological progression of user facts and preferences, often retrieving outdated or conflicting memories across sessions. Moreover, as conversation logs grow, exhaustive semantic search across unstructured history leads to prohibitive query latency and excessive token consumption. To address these limitations, we present SA-Mem, a temporal-centric memory indexing and topic-induced retrieval system designed for long-term LLM agents. SA-Mem introduces a Hierarchical Temporal Memory (HTM) architecture that organizes dialogue turns into semantically coherent MemBlocks, indexed by a bi-temporal interval tree that covers both session and event time. By employing a topic-induced retrieval pipeline, the system uses lightweight intent parsing to prune the search space via temporal filtering before performing fine-grained semantic ranking. Our demonstration shows that the temporal-centric indexing significantly enhances retrieval accuracy while reducing end-to-end latency. The demo also allows users to interactively compare retrieval with and without temporal awareness, inspect the internal HTM structure, and trace how retrieved memories affect answer generation.

Knowledge-intensive Web applications increasingly rely on Graph RAG to ground LLM answers in structured evidence, yet standard pipelines feed the reader a fixed top-k block of graph passages regardless of what each query requires, raising prompt-token cost and retrieval latency without a commensurate gain in accuracy. Analyzing this inefficiency at the passage-slot level on HotpotQA and 2WikiMultiHopQA, we identify two forms of sparsity: an oracle matches the recall of graph top10 with only about 5.5 passages on average, and only about 1% of the passages at graph ranks 6–20 recover gold evidence missing from graph top5. We propose MGEA (Marginal Graph Evidence Allocation), a lightweight allocator that keeps G5(q) as the base context and scores deeper graph passages by their conditional marginal gain over this context rather than by standalone relevance; it requires no extra LLM calls and is orthogonal to the underlying graph retriever. On 1,000 multi-hop QA examples, MGEA improves over graph top5 by +3.6 EM / +3.96 F1, outperforms budget-matched random, graph-score, text-only, MonoT5, and graph top7 baselines, and matches graph top8 while using fewer prompt tokens; the gains transfer across three readers and a second graph backend. A complementary dense-probe triage layer further cuts graph invocation to 52.7% of queries and retrieval latency from 6091 ms to 3002 ms while preserving high recall.

Technology-security risk analysis often turns on how policy, regulatory, and supply-chain events propagate over time. Topically relevant Web evidence is not enough. We formulate the task as query-driven reasoning over semantic event knowledge graphs. In this setting, typed event nodes, causal-temporal relations, policy actions, risk dimensions, temporal anchors, and provenance links support graph-grounded analysis. We propose a framework that constructs a query-aligned local event graph and selects a compact support trace. Trace selection uses risk-oriented scoring, candidate consensus, and boundary refinement, with query-intent completion used as a repair step. The system then generates answers with support-sensitive uncertainty assignment. The support trace is a first-class semantic object, not a post-hoc explanation. Experiments on TechRisk-EventLogic-Bench, a 120-query event-logic benchmark, show that although the framework does not maximize surface answer overlap, it improves support-trace faithfulness, boundary accuracy, grounded token alignment, and uncertainty matching over RAG baselines. Component-level ablations isolate support-trace candidate generation, query-intent rescoring, repair-style completion, and uncertainty control. Controlled public checks on MAVEN-ERE and ChronoQA further test event-relation support construction and temporal support ordering, without making leaderboard-style full-task claims. Overall, the results point to explicit support-trace selection as a useful mechanism for trustworthy graph-structured RAG over temporal event knowledge graphs.

We introduce Social Attribution QA Benchmark, a 1,200-item four-way multiple-choice benchmark for attribution-evidence-aware Social Attribution QA over FediData-derived Mastodon social Web data. Each item asks a system to choose the correct author or statement candidate from four options; the correct option is recoverable from observed source records and must satisfy the relevant authorship, temporal, and/or social-relation evidence required by the question. The benchmark covers two task families, WSW and WDWS, across eight subtypes, and includes evidence, constraint, and distractor-validation audit fields. As a reference solver, ATLAS implements constraint-first GraphRAG: it compiles question-implied constraints into graph filters before semantic ranking and comparative prompting. On Social Attribution QA Benchmark, ATLAS reduces 725,282 posts to a median predicted candidate pool of 3 posts and achieves 95.3% accuracy overall; SBERT-MiniLM with OracleFilter reaches 86.8% as the strongest diagnostic non-ATLAS configuration.

Fixed-context Web GraphRAG systems combine Web pages, graphs, and LLM readers, but operators still need to know which evidence condition preceded a suspicious answer. EvidenceFirst is an evidence-state monitoring and auditing framework for GraphRAG systems. Before generation, a deterministic checker assigns each question to a checked candidate path, repaired path, residual graph gap, or passage-fallback state; the saved state records the repair or fallback route, reader evidence, selector decision, and audit fields. The objective is evidence-state analysis rather than leaderboard optimization. Across HotpotQA-1000 and 2WikiMultiHopQA-500, EvidenceFirst remains within the operating range of recent GraphRAG baselines while exposing audit-risk queues that can be replayed without regeneration. Empirically, saved audit signals form review queues whose top-20% strata reach 67.0% and 40.0% error rates on HotpotQA and 2Wiki; adding graph-state fields to answer-selection signals improves error-detection AUC (0.6077 to 0.6321 on HotpotQA) and, more importantly, supplies typed routing causes: inside selection-conditioned queues, selected unrepaired-incomplete Hotpot cases and non-selected residual-gap 2Wiki cases are 26.6 and 33.3 error points worse than their respective comparison strata. Rather than predicting why a GraphRAG system failed, EvidenceFirst records a replayable evidence-state trace that helps failures be routed, audited, and recomputed after deployment.

While possessing strong general QA capabilities, Large Language Models (LLMs) frequently hallucinate on complex multi-hop KGQA tasks. Existing Graph-RAG methods attempt to incorporate structural knowledge but suffer from semantic-topological misalignment and information loss during naive graph linearization. We propose Structure-Guided Retrieval (SGR), a cascaded framework that bridges semantic matching and graph reasoning. SGR progressively executes semantic localization, entity-aware filtering, and topology-aware pruning to isolate high-confidence reasoning subgraphs. SGR converts retrieved graph paths into natural-language reasoning chains, bypassing intermediate topological noise. By decoupling retrieval from generation and adapting the model via Low-Rank Adaptation (LoRA), SGR leverages frozen base LLMs to perform transparent, knowledge-grounded inference without large-scale gradient updates. Evaluated on WebQSP and CWQ, SGR achieves 84.6% and 61.9% Hits@1, respectively. With an 8B-parameter open-source model (Llama 3.1) and a single forward pass, SGR achieves performance comparable to proprietary hundred-billion-parameter LLMs. SGR offers a robust and transparent solution for KGQA.

Long-term memory is essential for retrieval-augmented Web agents that support multi-session information seeking, personalization, and decision support. Existing memory systems often store conversation-derived facts, summaries, or chunks in a flat vector pool, which limits structured organization, cross-topic evidence aggregation, and interpretable multi-hop retrieval. This paper presents TopicMem, a topic-induced graph memory framework for long-term LLM agents. TopicMem segments conversations into topic-coherent spans, extracts structured multi-label memories, and organizes them with a dual graph index consisting of a hierarchical topic DAG and a topic association graph. During retrieval, TopicMem routes each query to candidate topics, ranks memories within selected topic pools, and expands to associated topics when multi-hop or causal evidence is required. Experiments on LoCoMo show that TopicMem improves retrieval over full-pool dense retrieval, increasing P@5 from 0.145 to 0.205 and R@5 from 0.372 to 0.457, corresponding to relative gains of 41.4% and 22.8%, respectively. In end-to-end question answering, TopicMem achieves 0.3640 F1 and 0.2840 BLEU-1, outperforming representative memory-enabled agent baselines. These results demonstrate that topic-induced graph memory provides an effective and interpretable foundation for long-term memory retrieval in Web agents.

Despite significant progress in the landscape of Vision-Language Models (VLMs), they remain susceptible to hallucination, wherein generated responses are not grounded in the visual input. External evidence can improve visual grounding; however, its effectiveness depends not only on the quality of the evidence but also on how the evidence is presented and how well its presentation aligns with a VLM’s evidence use characteristics. Consequently, the same evidence may correct a corrupt response in one VLM while corrupting an already corrected response in another. To address this issue, we propose VLM Evidence Aware Receptivity Profiling (VEARP) framework, an architecture agnostic and training free approach for characterizing how VLMs respond to external evidence using only observable inputs and outputs, thereby making it applicable to both open and proprietary black box VLMs. VEARP distinguishes evidence induced correction, corruption, preservation, and unresolved responses, and uses the resulting condition aware receptivity profile to derive model specific evidence use policies. We evaluate the framework in the context of object hallucination using two complementary object detectors, six VLMs (three open and three proprietary), and all three settings of the Polling Based Object Probing Evaluation (POPE) benchmark. The results reveal substantial model and condition dependent differences in evidence receptivity, demonstrating that the effect of external evidence on a VLM’s response is not determined by the evidence alone. Rather, the resulting response is a joint property of the model’s architecture, the evidence presentation strategy, and the reliability of the provided evidence.

GenVis is an interactive system for exploring the generation of synthetic genotype data under privacy-utility constraints. Building on prior work on adaptive optimization for synthetic genotype generation, the system exposes the rollback-enabled loop used to refine synthetic datasets through biological utility metrics and membership-inference-based privacy analysis. It supports both on-demand generation and the inspection of precomputed runs, allowing users to follow model trajectories, rollback events, and privacy–utility dynamics without rerunning the entire pipeline. Beyond serving as a data generation tool, GenVis provides visualizations of genetic distances, allele frequencies, and linkage disequilibrium patterns, enabling detailed structural comparison between real and synthetic data and showing how privacy constraints affect the generator’s ability to preserve genomic patterns. We conclude with a discussion of system behavior, design limitations, and practical considerations that arise when applying adaptive synthetic-genomics pipelines in interactive settings.

Smart agriculture increasingly relies on IoT-enabled systems that generate large-scale, sensitive, farm-specific data to enable intelligent agricultural decision-making. However, these systems must ensure strong privacy preservation while handling highly heterogeneous, non-IID data distributions. Existing approaches are typically static, with limited adaptability to heterogeneous and evolving conditions, which constrains their performance in dynamic agricultural environments. This paper proposes the Adaptive Hybrid Federated Meta-Recommender System (AHFMRS), a decentralized federated framework that integrates federated learning, heterogeneous recommendation models, and a runtime meta-decision layer to enable privacy-preserving and context-aware decision-making in distributed IoT-based agricultural environments. The framework is evaluated on three heterogeneous, multi-domain datasets encompassing time-series agricultural data and UAV-based edge observations. Experimental results over prolonged federated training (50 rounds) demonstrate that AHFMRS achieves a peak accuracy of 66% and an F1-score of 65.3%, consistently outperforming baseline architectures while provid ing an efficient and robust decision-making solution under diverse and resource-constrained agricultural environments.

Digital Responsibility (DR) highlights the principles that should guide all aspects of digital technologies’ usage. The expansion of digital technologies across different professional domains gave rise to DR requirements related, among else, to data privacy, transparency, fairness and sustainability. The motivation of this study was to assess the association of DR skills with occupations, using as an example the Computer Science, Business Admin-istration and Educational Psychology domains. ESCOX, an LLM-based tool that capitalized on the ESCO classification, was employed to extract the frequency of DR skills related to occupations across these three do-mains, based on Online Job Advertisements that were retrieved from multi-lingual EU digital sources. Hierarchical clustering and network analysis were employed to explore the association of DR skills and occupations within each domain. The findings of the study showed distinct DR skill profiles across the three domains, with Data Privacy related skills exhibit-ing importance in all three domains. This study provides a supporting tool for individuals regarding the DR skillset required to successfully apply for an occupation within the studied domains and may set the basis for further research that will cover the whole labour-market landscape.

The paper investigates the possibility of using LLMs for automatized assessment of museum accessibility based on the relevant content published at the respective museum websites. Results of experimental research performed on the websites of 20 most popular museums in the world are reported which regard the correctness of LLM-based assessment, similarity of assessment by different LLMs, and the capability of LLMs to review the assessment results produced by other LLMs. The study conclusions open the way for improvements in the practice of museum accessibility assessment.

As hotels increasingly invest in cultural elements through architectural styles, spatial decoration, and regional cultural elements, assessing whether such cultural investments can translate into guest attraction has become an important managerial issue. However, user feedback is distributed across review texts and hotel images, and the two modalities contribute unequally to identifying culture‑driven accommodation decisions. Moreover, real‑world data often contain missing modalities. To address these challenges, this paper proposes Text‑Dominant and Visual Posterior Calibration (TVPC) framework for hotel cultural attraction identification. It first introduces a Corpus‑Driven Cultural Semantic Evidence Aggregation (CCSEA) module to capture culture‑related evidence from review texts through unsupervised semantic‑cluster mining and cluster‑guided attention. It then employs a Cultural Visual Evidence Interface (CVEI) to extract cultural visual support from hotel images. Furthermore, a Fusing Support Strength, Consistency and Confidence (FSCC) module is designed to incorporate visual evidence as a weak posterior correction signal by jointly modeling visual support strength, image‑text consistency, and textual confidence under a text‑dominant decision framework. Finally, a modality‑aware routing mechanism enables unified processing across multimodal and unimodal input scenarios. Experimental results demonstrate that TVPC achieves an F1‑score of 0.7091, improving over the CLIP ViT‑B/32 baseline by 8.73 percentage points. Moreover, TVPC maintains competitive performance on four generalization tasks, achieving F1‑scores between 0.7012 and 0.8031, which validates its effectiveness and structural adaptability.

]>

In graph-based recommender models, the learning signals involved in personalized ranking may differ in their relevance and informativeness, observed positive interactions may exhibit different consistency with a user’s recent interests, and uniformly sampled negatives can provide insufffcient ranking difffculty. To address these issues, we propose TMR-Rankformer, a multi-level signal reffnement framework built upon the ranking-inspired Rankformer backbone. At the node level, TMRRankformer constructs a recency-aware preference prototype from each user’s recent training interactions and uses it as an auxiliary correction to the user representation while retaining the collaborative information encoded by the complete interaction graph. At the positive-interaction level, the prototype is used to measure the consistency between an observed positive item and the user’s recent preference, and the resulting bounded weight adaptively reweights the pairwise ranking loss. At the negative-sample level, a mixed hard negative sampling strategy mines an informative negative from a randomly sampled candidate pool and probabilistically combines hard-negative and conventional random-negative supervision. Experimental results on Ali-Display, Epinions, Amazon-CDs, and Yelp show that TMR-Rankformer consistently outperforms Rankformer and other baselines on NDCG@20 and Recall@20. Compared with Rankformer, TMR-Rankformer achieves NDCG@20 improvements of 2.61%, 2.53%, 1.56%, and 1.45% on the four datasets, respectively. Ablation and hyper-parameter analyses further verify the complementary contributions of the three reffnement components.

This study compares three audio deepfake detectors under cross-domain shift: a Siamese CNN trained with StacLoss, AASIST on raw waveforms, and VDD using fused STFT/MFCC/CQCC features. The Siamese and VDD systems are trained on ASVspoof 2019 LA, while AASIST uses the authors' official pre-trained weights. On ASVspoof 2021 DF, the Siamese model performs best (21.59% EER; 85.53% ROC-AUC), ahead of AASIST (34.09%; 72.20%) and VDD (39.42%; 66.70%). On In-the-Wild, all three approach chance-level detection (43.02–44.45% EER). The results identify training-distribution mismatch and score calibration as major barriers to real-world robustness and suggest that training objective may influence transfer as strongly as architecture.

Caregivers hold almost all of a child's waking hours; specialists hold one to three a week. The World Health Organization (WHO) Caregiver Skills Training (CST) program was designed against that asymmetry, and the agentic capability to operate such a program is now purchasable by subscription. Yet no such system exists. The received explanation is that health data is locked inside institutions. That explanation is incomplete. Practice management systems record what a child did. They do not record why a clinician changed the program, and session notes carry that reasoning only as prose never bound to the change it explains. Unlocking every silo would not turn it into a corpus. We separate three data problems the field conflates, training, retrieval and evaluation, and show only the first is solved by procurement. We then give a reference architecture in which the corpus is a byproduct of a service useful on day one: documentation relief needs no ground truth and emits decision-annotated records as exhaust. We also give an annotation schema, a mapping from the five CST domains onto supervised agent roles with their failure modes, and a cost model in which continuous video exceeds a family-affordable price roughly sevenfold while event-triggered capture consumes a few percent of it.

Counterfactual explanations (CE) provide intuitive insights into machine learning predictions by identifying input modifications that lead to a desired outcome. While widely studied for classification, extending them to regression remains challenging due to the continuous nature of the target variable and the need to preserve semantic plausibility. Existing methods typically rely on feature-wise perturbations and often overlook variable interactions that are prevalent in real-world data, which can lead to counterfactuals that fail to respect structural dependencies learned by the model. We propose SIGCE, a novel framework for generating interaction-aware counterfactual explanations for regression. SIGCE combines Shapley interaction values with groups of highly related variables derived from hierarchical clustering to identify influential interaction-aware feature subsets and constrain the search space accordingly. This design encourages coordinated feature updates and yields counterfactuals that are sparser and closer to the original instance. Experiments on multiple regression datasets show that SIGCE delivers the best overall trade-off between proximity, sparsity, and success rate compared with representative baselines, yielding counterfactual explanations that better reflect the interaction structure captured by the predictive model.

Temporal graph neural networks have demonstrated strong performance in dynamic link prediction, yet understanding the factors behind their predictions remains a challenging task. Most existing methods rely on post-hoc explainers that operate separately from the predictive model, potentially leading to inconsistencies between prediction and explanation. To address this limitation, we introduce Dyn-CFX, a counterfactual explanation framework based on a single Attention-based Dynamic Variational Graph Autoencoder (ATT-dynVGAE), jointly used for both prediction and explanation. By exploiting the same variational encoder–decoder architecture, Dyn-CFX performs counterfactual reasoning directly within the learned latent space, avoiding predictor–explainer mismatch and eliminating the need for retraining. For a predicted link, Dyn-CFX evaluates the effect of historical edge removals through structural interventions and quantifies their contribution using a temporally weighted conditional marginal score. Experiments conducted on five realworld temporal networks show that the proposed framework produces explanations that are more faithful and temporally consistent than those of existing baselines, while maintaining competitive predictive accuracy

Drug--target interaction (DTI) and affinity (DTA) prediction are central to computer-aided drug discovery, and deep models based on graph neural networks, transformers, and protein language models now achieve strong accuracy on benchmarks such as Davis, KIBA, BindingDB, and PDBbind. Accuracy alone, however, is insufficient when chemists need to understand which molecular fragments, residues, pockets, or intermolecular contacts support a prediction. Most existing approaches rely on post-hoc explanations, which are generated after training and may not faithfully reflect the model’s decision process. This survey reviews DTI/DTA and binding-prediction methods through the lens of explanation-guided learning (XAI-GL), where explanation signals influence model training rather than being assessed only after prediction. We argue that drug-target prediction is a particularly suitable testbed for XAI-GL because explanation targets can often be grounded in physical evidence, including non-covalent interactions, contact maps, and binding pockets. We map representative models, including MONN, ArkDTA, and DeepRelations, onto general XAI-GL mechanisms such as auxiliary explanation supervision, explanation regularisation, and interpretable bottlenecks. We further propose a two-axis taxonomy organised by explanation substrate and by the role of explanation in the learning pipeline. This taxonomy shows that the literature remains concentrated around post-hoc attribution and unsupervised attention, while supervised, faithfulness-oriented explanation-guided learning remains underexplored. Finally, we identify open challenges around standardised benchmarks, the faithfulness-plausibility gap, explanation leakage, partial and noisy explanation labels, and generalisation to unseen drugs and targets.

Streaming anomalous subsequence detection aims to contin- uously identify abnormal temporal patterns from streaming time series and plays a critical role in numerous real-world applications, including industrial monitoring and Internet-of-Things systems. Although the ex- isting methods have already made progress, they face three fundamental challenges: (1) inability to handle concept drift; (2) lack of an online update mechanism; and (3) threshold sensitivity. In this paper, we pro- pose RL-OSAD, a novel reinforcement learning framework for stream- ing anomalous subsequence detection. Specifically, RL-OSAD integrates adaptive normal-pattern modeling and sequential decision-making into a unified architecture. It first employs a Sequence Data Clustering Module (SDCM) to maintain a weighted set of normal-shape prototypes that are adaptively updated to capture evolving normal patterns in data streams. The prototype-aware anomaly score generated by SDCM is then incor- porated into an Anomalous Subsequence Detection Module (ASDM), which formulates anomaly detection as a Markov decision process and learns a stochastic policy. Extensive experiments on multiple real-world datasets demonstrate that RL-OSAD consistently outperforms state-of- the-art static and streaming anomaly detection methods in terms of de- tection accuracy and robustness.

Deep drug-target affinity (DTA) predictors can be accurate yet opaque, and the dominant response in the field has been to explain a trained model after the fact. This separation between prediction and explanation leaves faithfulness largely assumed. Explanation-guided learning offers a different route: explanation quality becomes part of the optimisation objective. We present XGL-DTA, a self-explaining DTA model that encodes a drug as both a molecular graph and a SMILES sequence, encodes the protein as a residue sequence, learns soft rationale masks over atoms, tokens, and residues, and trains these masks with sufficiency, comprehensiveness, sparsity, consistency, and stability objectives added to the affinity-regression loss. We evaluate the model in three complementary experimental scenarios. First, internally generated five-fold splits test generalisation under random, cold-drug, cold-target, cold-both, and cross-dataset conditions. Second, a controlled external-split protocol reuses the verifiable DeepDTA Davis folds and GraphDTA KIBA folds to compare the full self-explaining model with an otherwise identical no-explanation ablation. Third, a global explanation-weight scale traces the accuracy-faithfulness frontier. On the external folds, paired Wilcoxon tests find no statistically detectable difference between the full and no-explanation variants for either MSE or concordance index, with all p-values greater than 0.18. On Davis, increasing the explanation scale from s = 0 to s = 0.5 reduces sufficiency from 0.224 to 0.013 while leaving MSE essentially unchanged, 0.281 versus 0.282. These results frame XGL-DTA as an accuracy-faithfulness contribution rather than a raw-performance claim: faithful self-rationales become a controllable property of DTA prediction without a measurable accuracy cost under the studied protocol.

Unmanned aerial vehicles (UAVs) equipped with onboard sensors have been widely deployed in various applications. Recently, a new paradigm called UAV crowdsensing (UCS) has emerged. UCS servicing delay-sensitive web applications has drawn rising attention by navigating UAVs to perform sensing tasks within a specified deadline to ensure data freshness, which is crucial for real-time applications such as disaster response, emergency monitoring and live webcast. In this paper, we focus on multi-objective UAV crowdsensing (MOUCS), where achieving high data quality necessitates complete data collection from all PoIs, and fair data collection must be taken into account. Specifically, we investigate the utility-fairness trade-off in delay-sensitive UCS and introduce two metrics, namely Utility Ratio and Fairness Index, to evaluate the data quality and geographic fairness of sensing tasks, respectively. A meta-learning enhanced multi-objective deep reinforcement learning (MORL) approach called “Meta-MOUCS” is designed to learn a generalized policy to solve the problem. It comprises a distributed learner-trainer architecture integrated with spatial-temporal modeling for multi-task division explicitly, and a meta-learning mechanism for fast adaptation. Simulation results show that Meta-MOUCS outperforms two baselines in HV and PF, demonstrating its effectiveness in addressing the utility-fairness trade-off in delay-sensitive UAV crowdsensing.

Companies accumulate large volumes of textual data from customer interactions. When sufficiently granular and reliably labeled, this data can reveal customer needs and pain points, helping guide service improvements. Manual labeling, however, is costly and error-prone. This paper investigates how well different modelling approaches classify call center conversations into a two-level hierarchy of topics and subtopics. We compare three families of models: (i) traditional NLP baselines, namely a hierarchical Support Vector Machine (SVM) and a neural network method (HDLTex++), both trained on text embeddings; and (ii) several off-the-shelf Large Language Models (LLMs). The LLMs are deployed in a flat configuration, where topics and subtopics are predicted jointly, and in a two-step hierarchical configuration that separates level-1 and level-2 decisions. We evaluate GPT-5-mini, GPT-5-nano, Grok-4-fast-non-reasoning, and Grok-3-mini using macro- and micro-averaged hierarchical precision, recall, and F1 on the joint label space. The LLMs consistently outperform the SVM and HDLTex++, and we report a ranking of the individual LLMs across configurations. The hierarchical LLM scheme reduces both accuracy and efficiency relative to the flat variant. An additional efficiency analysis highlights a clear accuracy–runtime trade-off: SVM and HDLTex++ are faster but substantially less accurate than the LLMs. Overall, our findings show that off-the-shelf LLMs deliver superior hierarchical multi-label classification performance to traditional NLP techniques, and that a flat prediction scheme is preferable to a hierarchical one for relatively simple taxonomies and tasks.

Knowledge tracing (KT) aims to model students’ knowl edge states from historical learning interactions and predict their future performance. Conventional deep learning (DL)-based KT models often lack pedagogical interpretability, while large language model (LLM) based KT methods can generate natural language explanations but still lack stable educational evidence, which may lead to unstable judgments. To address these limitations, we propose PEAR-KT, a peer-experience augmented LLM framework for knowledge tracing that leverages representative peer students as external evidence to support more reliable and interpretable prediction. First, we construct a Cognitive Memory Bank (CMB) to preserve representative peers’ learning trajectories, knowledge states, and structured cognitive attributions, thereby providing inter pretable peer references for the target student. Meanwhile, we build a Prediction Experience Bank (PEB) to summarize historical prediction contexts, model judgments, ground-truth outcomes, and error attributions, thus offering experience-guided constraints for bias-aware adjust ment. During online prediction, PEAR-KT retrieves cognitively similar peer cases from the CMB and matches relevant experiences from the PEB through hierarchical retrieval. The retrieved evidence, together with the target student state and question information, is organized into structured prompts to guide a frozen LLM in generating the final prediction and explanation. Experiments on ASSIST12, Eedi, and DBE-KT22 demonstrate that PEAR-KT achieves competitive or superior performance over representative baselines, validating the effectiveness of peerexperience-augmented reasoning for explainable knowledge tracing.

Federated learning enables collaborative model training without sharing raw data, but it still faces two key risks: reliance on a central coordinator and vulnerability to poisoned model updates. Blockchain-based federated learning can reduce centralized trust, yet blockchain records alone cannot determine whether submitted updates are reliable or explain validation decisions. To address this gap, this paper proposes RABFL-CV, a robust and auditable blockchain federated learning framework with committee validation. In each round, a reputation-aware validator committee evaluates submitted model updates and produces signed evidence, while a selected miner aggregates committee votes, updates reputation, applies stability-gated aggregation, and records the decision evidence on chain. Experimental results show that RABFL-CV maintains training continuity under single-point failures, achieves stable accuracy under multiple model-poisoning attacks, and supports long-term tracking of malicious participants through reputation evolution. We also implement and open-source a Ray-based prototype that abstracts key blockchain-FL roles, committee behaviors, and security parameters, providing a configurable testbed for designing and evaluating security mechanisms. The code is publicly available at: https://anonymous.4open.science/r/RABFL-CV-D6A3/.

Roadside OCR remains unreliable when character strokes are incomplete, corrupted, or visually misleading. This issue is particularly severe for license plate recognition for muck trucks, where mud adhesion, paint erosion, glare, low illumination, perspective deformation, and non-standard layouts may erase genuine strokes while introducing stroke-like artifacts. Existing restoration-based recognizers and multimodal pipelines can generate plausible plate sequences, but they often lack explicit verification of whether the predicted characters are supported by visible stroke topology. To address this problem, we propose TopoGuard, a topology-constrained and degradation-aware framework for this task. TopoGuard mitigates structural hallucination by learning stroke-aware representations with topology destruction-driven adversarial masking and by performing constraint-guided decoding based on layout validity and visible topology consistency. To handle non-stationary degradation, it further introduces a scene-prior gated MoE-LoRA module, which routes samples with heterogeneous degradations to specialized low-rank experts and reduces negative transfer across low-light, overexposed, deformed, and occluded scenarios. Experiments on a self-built muck truck dataset and CCPD subsets show that TopoGuard improves recognition accuracy, reduces the structural error rate, and maintains practical RoI-level inference efficiency under regular, illumination-degraded, and geometrically distorted conditions. Code is available.

As microservice systems become increasingly complex, root cause localization (RCL) plays a critical role in identifying the underlying causes of system faults. Multi-modal monitoring data provide complementary perspectives for characterizing system faults, and many works propose to extract events from them for RCL. However, utilizing multi-modal monitoring data for RCL faces challenges: multi-modal monitoring data are heterogeneous in granularity, and many works fail in fine-grained modeling of event propagation and its influence on root cause. To address these issues, we propose MCEG-RCL, an RCL framework based on a continuous-time dynamic event graph. MCEG-RCL converts multi-modal monitoring data into unified events and builds a multi-modal continuous-time dynamic event graph, including metric, log, and trace nodes. The framework leverages neural ODEs to capture anomaly state evolution between irregularly spaced events and employs GRUs to model state jumps when events arrive. In this manner, the framework generates robust anomaly states that are informative for root cause localization. Experimental results on three benchmarks demonstrate that MCEG-RCL achieves accurate and efficient root cause localization while exhibiting strong robustness and generalization.

Selecting execution plan for concurrent queries is challenging in database management systems, because the interactions between queries (e.g., resource contention and sharing) are highly complex and change dynamically with the execution of operations. The key to selecting execution plans for concurrent queries is to measure query interactions (QIs) accurately. The state-of-the-art methods consider the usage of system resources at the moment when a query is executed while ignoring the change in usage during the execution of queries, which introduces errors to the measurement. To address this issue, we propose a time series heterogeneous graph for describing the dynamics of QIs. The model incorporates the temporal dimension to describe the dynamic changes in QIs and an evaluation model for calculating the weights of the edges between operations at any timestamp. A time series heterogeneous graph classification model is proposed to make use of QIs dynamics, in which a long short-term memory neural network is used to learn temporal relationships from multiple temporal graph representations and select the appropriate execution plan for queries. Experiments on PostgreSQL demonstrate that the average accuracy of our model is 51.2% higher than that of the query optimizer and 6.7% higher than that of the state-of-the-art models.

Programming knowledge tracing (PKT) estimates learners’ programming knowledge states from historical programming interactions for future performance prediction. Existing PKT approaches primarily focus on response results, often neglecting the rich information contained in students’ iterative code submissions and the semantic differences among learning behaviors under different time intervals, yet these factors are crucial for accurately estimating learners’ evolving knowledge states in programming tasks. To address these limitations, we propose (SETR-PKT), a Strategy Evolution Graph and Time-Aware Retrieval framework for programming knowledge tracing. Specifically,by analyzing students’ consecutive code submissions on the same programming question, we construct a strategy evolution graph to characterize the evolution of their problem-solving strategies during iterative debugging and solution refinement. Next, to accurately characterize students’ knowledge states, we introduce a knowledge state encoder that integrates multiple learning sources, including strategy representations and interaction behaviors. Finally, to distinguish learning behaviors under different time intervals, we develop a dual-time knowledge retrieval decoder to retrieve historical knowledge when answering future exercises. Extensive experiments on three real-world datasets demonstrate that the proposed model outperforms state-of-the-art methods, achieving average improvements of 6.06% in AUC and 5.87% in ACC within baselines.

Knowledge graphs provide structured evidence for large language models, yet automatically constructed domain graphs often contain redundant triples and ambiguous entities. Redundant triples enlarge the retrieval space, increase storage cost, and slow down graph-based retrieval-augmented generation. Ambiguous entities weaken evidence quality because specialized concepts may lack explicit definitions. We propose \textbf{DELG}, a framework that constructs definition-enriched lightweight domain knowledge graphs. DELG first extracts triples from raw domain texts, then removes semantic and structural redundancy through semantic clustering, functional dependency checking, bidirectional normalization, and transitive redundancy detection. DELG further computes entity complexity from information amount and domain specialization, and recursively enriches complex entities with simpler definitional triples. Experiments on a medical corpus from \textit{Treating Autoimmune Disease with Chinese Medicine} show that DELG improves retrieval recall, answer accuracy, and inference latency in LLM-based KG-QA. Compared with a full KG, the redundancy-reduced graph lowers average inference latency by 20.87\%, while recursive definition enrichment gives the best overall accuracy under GPT-4o-mini and improves recall under a locally deployable Qwen2.5-7B-Instruct setting.

Large Language Models (LLMs) have demonstrated their remarkable capabilities in document understanding. However, recent research reveals that LLMs still exhibit performance gaps in Document-level Relation Extraction (DocRE) as requiring fine-grained comprehension. The commonly adopted "extract entities then predict relations" paradigm in LLM-based methods leads to these gaps due to two main reasons: (1) Numerous unrelated entity pairs introduce noise and interfere with the relation prediction for truly related entity pairs. (2) Although LLMs have identified semantic associations between entities, relation labels beyond the predefined set are still treated as prediction errors. To address these challenges, we propose a novel "Relation as a Prior"(RelPrior) paradigm for LLM-based DocRE. For challenge (1), RelPrior utilizes binary relation as a prior to extract and determine whether two entities are correlated, thereby filtering out irrelevant entity pairs and reducing prediction noise. For challenge (2), RelPrior utilizes predefined relation as a prior to match entities for triples extraction instead of directly predicting relation. Thus, it avoids misjudgment caused by strict predefined relation labeling. Extensive experiments on two benchmarks demonstrate that RelPrior achieves state-of-the-art performance, surpassing existing LLM-based methods.

Next Point-of-Interest recommendation forecasts a user's future destination based on historical trajectories, vital for human mobility modeling in Location-Based Social Networks. Existing methods have achieved notable progress by capturing sequential dependencies and spatio-temporal contexts. However, dominant chronological methods inevitably entangle periodic routines with stochastic noise, rendering mobility patterns highly vulnerable to perturbations. Moreover, existing spectral attempts typically model trajectories in isolation, leaving the coupling of frequency-enhanced signals with heterogeneous contexts a critical bottleneck. To address these limitations, we propose FPST-POI, a Frequency-enhanced Personalized Spatio-Temporal Multi-View Fusion framework for next-POI recommendation. Specifically, FPST-POI first develops a Frequency Enhancer that adaptively balancing low-frequency periodic patterns and high-frequency behavioral variations. To integrate heterogeneous contexts, we design a Multi-view Gated Fusion module that adaptively aligns and aggregates the User-POI view and the Time-Region view. The fused check-in representations are then fed into a Transformer encoder and MLP decoding. Extensive experiments on real-world check-in datasets demonstrate that FPST-POI consistently outperforms competitive baselines.

Multimodal Knowledge Extraction (MKE) is a vital research branch of multimodal knowledge graphs (MKGs), focusing on extracting entities and relations from multimodal data. Although existing studies can perceive and integrate data representations, there are still problems, such as high fine-tuning costs and high risk of catastrophic forgetting. In this paper, we propose a Cognitively-inspired Parameter-Efficient MKE model, namely CogPEMKE. Specifically, we draw on the rapid adaptation theory and design a low-rank adapter for calculating the attention matrix, which efficiently adapts downstream tasks with minimal parameters. Additionally, we simulate the sparse encoding mechanism of the brain and present a sparsely activated attention mechanism by taking advantage of the "divide and conquer" feature of mixture-of-experts networks, which effectively enhances the model's capacity and expressive power, but only some parameters are activated. Extensive experimental results on MKGs demonstrate that the CogPEMKE can achieve the state-of-the-art (SOTA) performance compared to current MKE models.

Cross-document event coreference resolution aims to determine whether event mentions across multiple documents refer to the same real-world event. In practice, however, applications such as news aggregation and event monitoring require processing event mentions as they arrive, making online event coreference a practical necessity. Moreover, multi-modal information can provide additional evidence when textual descriptions alone are insufficient. In this paper, we solve the task of streaming multi-modal event coreference resolution across documents, where models receive temporally ordered text-image pairs and must make irreversible create-or-link decisions based solely on historical evidence. This task faces two challenges: (1) delayed evidence may arrive only after decisions have already been made, and (2) textual descriptions often share similar surface patterns, necessitating visual signals for disambiguation. Existing methods assume offline scenarios with complete document collections and lack benchmarks for our task. To fill this gap, we first construct two new benchmarks: ECB-OL and WCE-OL. Next, we propose PSMEC, a streaming coreference resolution framework that maintains persistent event records accumulating multi-modal evidence from historical observations. We propose PSMEC—a streaming coreference framework using persistent event records to accumulate multi-modal evidence. Its shared encoder outputs both create/link scores and a false-link risk adjustment to refine scores, with incremental record updates after each commit. Experimental results demonstrate that PSMEC outperforms baseline methods on both benchmarks that we construct.

Cross-modal Hashing (CMH) shows great potential for large-scale multimodal retrieval, but existing methods face two primary bottlenecks: (1) failing to bridge the gap between coarse-grained labels and high-dimensional distributions while ignoring underlying semantic topology; and (2) struggling to disentangle shared semantics from modality-private features, resulting in a lack of latent space discriminability. To address this, we propose Semantic-Enhanced Dual-Prototype Decomposition for Cross-Modal Hashing (SDDH). SDDH introduces dual learnable prototypes to establish structured semantic anchors in both feature and latent spaces. It leverages Pointwise Mutual Information (PMI) to enhance latent prototypes for capturing deep categorical topology. Additionally, a prototype-guided disentanglement mechanism uses strict orthogonal constraints to strip away private features, successfully anchoring the purified semantic core to the enhanced prototypes. Extensive experiments across three datasets validate the superior performance of SDDH.

In this paper, we introduce AREF-RAG, an Agentic Refinement Framework for Retrieval-Augmented Generation. AREF-RAG uses retrieval feedback during index construction to improve chunk representations before final indexing. The method first generates page-level probe questions, creates an initial chunk set, stores it in a temporary vector store, and evaluates whether retrieval returns chunks from the expected source pages. When retrieval performance does not meet the acceptance criteria, an evaluator agent analyzes the failed cases and produces page-level rewriting guidance. A rewriter agent then updates only the failed chunks while preserving their original structure. The refinement loop is repeated until the retrieval criteria are met or the maximum number of iterations is reached. Experiments on the open-source FinanceBench release show that AREF-RAG improves Recall@10 by 22.92% relative to the best baseline, increasing from 0.3404 with semantic chunking to 0.4184. Instead of treating retrieval only as a downstream evaluation step, AREF-RAG uses retrieval failures as feedback for improving the indexed document representation. The goal is to make chunks more self-contained and easier to retrieve, leading to better evidence recovery for downstream question answering.

Learning to hash compresses high-dimensional data into compact binary codes for sub-linear approximate nearest-neighbour search. Yet almost all deep hashing methods assume a closed, static class space, whereas real retrieval systems face an open world in which new categories arrive continually. Updating the hash function for new classes triggers catastrophic forgetting and, worse, forces a costly re-encoding of the entire gallery. We trace this failure to two coupled sources of drift: the per-class hash centres move as new classes are added (\emph{center drift}), and the backbone representation shifts during incremental updates (\emph{feature drift}). We observe that a frozen vision--language model (VLM) supplies a training-invariant semantic coordinate system: the text embedding of a class name is a deterministic constant, independent of which classes have been seen or how long the model has trained. Building on this, we propose \emph{Language-Anchored Continual Hashing} (\method), which (i) defines every class's hash centre as the binarised projection of its frozen text embedding and locks it forever, (ii) freezes the VLM image encoder so features never drift, and (iii) preserves old-class codes without storing old samples through a text-proxy anti-forgetting objective. Because the centres are text-derived, a new class is assigned a valid centre before any of its images are seen, and old gallery codes are never rewritten. The text-proxy objective keeps later queries compatible with those stored codes, so old codes stay usable without re-indexing. We evaluate on class-incremental retrieval benchmarks against pre-VLM incremental hashing, training-free VLM baselines, and joint-training oracles.

Retrieval-augmented generation (RAG) has become an effective way to improve large language models (LLMs) for vertical-domain question answering (QA) by grounding generation in domain knowledge. However, complex document QA in vertical domains remains challenging for existing RAG methods, which still struggle to jointly support fine-grained multimodal query understanding and coordinated multi-stage retrieval. As a result, they often produce less effective answers, require longer response times, and consume more tokens during retrieval and generation. The root cause is that complex query intentions are not modeled with sufficient fidelity: current methods cannot reliably identify the structured constraints and unstructured knowledge needs embedded in a query, nor convert them into executable operations that directly guide retrieval, evidence organization, and answer generation. To address this issue, we propose HUQI-RAG, an agentic RAG framework for complex document QA guided by holistic understanding of query intention. HUQI-RAG maps heterogeneous query intentions to fine-grained multimodal query types composed of query modes and knowledge needs, and coordinates retrieval, evidence refinement, and answer verification over an MMO knowledge space, forming a multi-round closed loop with adaptive termination. Experiments on complex document QA tasks across multiple vertical domains show that HUQI-RAG achieves strong overall performance while remaining efficient in response time and token cost.

Multi-vector retrieval improves fine-grained semantic matching but incurs high storage and computation cost. For large-scale long-document collections, keeping all full-precision token embeddings resident in memory is often impractical. Therefore, under the ColBERTv2 compression scheme, existing systems usually adopt a two-stage pipeline with candidate generation followed by candidate refinement. Although recent studies have greatly accelerated candidate generation, candidate refinement remains expensive because it still needs to decode and score large numbers of compressed document tokens. We observe that only a small fraction of document tokens contributes to the final MaxSim score. Based on this observation, we propose CoLT, a lightweight compression-based token selection method for candidate refinement. CoLT first computes approximate inner products between query tokens and non-redundant cluster centroids, then selects a small set of promising document tokens for each query token, and finally performs deduplicated decoding before exact scoring. Experiments on Quora, MS MARCO, and LoTTE pooled show that, after integration with IGP, CoLT substantially reduces decoded document tokens and improves the latency-recall trade-off. In long-document settings, CoLT accelerates candidate refinement by about 5x and lowers end-to-end retrieval latency by around 2x while maintaining competitive recall.

Cold-start multimodal recommendation faces three coupled challenges: (i) sparse interactions obscure user intent, (ii) cold items remain topologically isolated, and (iii) similarity-based item graphs may cause semantic drift. To address these issues, we propose MOTIF, a Motivation-guided Topology Inference framework for cold-start multimodal recommendation. MOTIF integrates Semantic Motivation Reasoning, Knowledge-enhanced Graph Reconstruction, Weighted Graph Contrastive Learning, and Semantic-Structural Alignment. It uses offline LLM reasoning to infer motivation semantics, reconstructs transferable item-item topology, and learns robust graph embeddings without injecting generated text into prediction. Experiments on three multimodal benchmarks show consistent gains over graph-based, multimodal, cold-start, and LLM-enhanced baselines, with up to 6.07% relative improvement over the strongest recent baseline.

Parametric retrieval-augmented generation (PRAG) alleviates the inherent inefficiency and noise sensitivity of RAG by internalizing retrieved documents into model parameters, typically in the form of LoRA. However, the linear combination of multiple LoRAs can easily lead to parameter conflicts and geometric perturbations, impairing the capabilities of large language models (LLMs). Orthogonal fine-tuning provides inspiration for stable fusion. Nevertheless, orthogonality constraints increase training complexity, and multi-matrix fusion faces a dilemma: linear averaging destroys orthogonality, whereas the Riemannian geometric mean is computationally prohibitive. Moreover, orthogonal transformations alone struggle to express scaling variations and non-orthogonal knowledge, making them unsuitable as a direct replacement for LoRA. To address these issues, we propose ORDPRAG, a geometry preserving orthogonal-residual decoupling framework for PRAG. Specifically, we decompose each LoRA into a structure-preserving orthogonal transformation and a flexible additive residual through orthogonal-residual decoupling, thereby mitigating geometric perturbations and parameter conflicts. By mapping orthogonal transformations into the Lie algebra $\mathfrak{so}(d)$, we perform magnitude-corrected fusion to adaptively handle both direction and strength, thus resolving the matrix fusion challenge. The residual components that cannot be explained by orthogonal transformations are fused in Euclidean space. In this way, ORDPRAG preserve the geometric structure of the base weights of LLMs while retaining the ability to inject new knowledge. Experiments show that ORDPRAG significantly alleviates the geometric perturbations caused by linear combinations of LoRA and improves the reasoning performance of LLMs.

Efficient vector search over massive datasets has become a key infrastructure component for retrieval-augmented generation (RAG) and multimodal systems. Although memory-based indexes offer low query latency, their high cost and limited capacity make them difficult to sustain at the billion scale and beyond. As a result, approximate nearest neighbor search is increasingly shifting toward disk-based indexing, with disk-based graph indexes often standing out for their strong query throughput. Yet two issues remain in disk-based graph search. First, the logical locality of graph indexes does not naturally map onto storage locality, resulting in poor I/O efficiency and many wasteful reads during query expansion. Second, the access window during the I/O-intensive expansion phase shifts continuously, making it difficult for subsequent queries to consistently reuse previously loaded regions. To address these issues, we propose a collaborative optimization framework for disk-based graph ANNS that combines offline storage layout optimization with runtime caching. The former improves the alignment between logical and storage locality in graph indexes, while the latter introduces a dynamic caching mechanism tailored to the I/O-intensive expansion phase. Overall, the framework reduces wasteful disk I/O during query expansion and improves cache reuse during the I/O-intensive expansion phase, leading to more efficient query processing for disk-based graph indexes. Experiments on three 100M-scale datasets show that the proposed framework consistently outperforms existing SOTA methods, reaching a peak 2.13x QPS gain and still delivering 1.07x-1.56x the QPS of SOTA methods at the main operating point, while reducing mean I/Os to 0.77x-0.97x.

Multimodal reasoning has become increasingly important in industry-scale applications that require joint understanding of text, tables, charts, figures, and layout-rich visual content. One particularly challenging setting arises when evidence is distributed across hundreds of pages and multiple modalities. Existing Multimodal Large Language Models (MLLMs) address this problem either through flat long-context processing, which produces prohibitive context cost and structural information loss, or through retrieval-augmented pipelines that depend on fixed granularity retrieval and often generate ungrounded outputs. Motivated by these observations, we introduce HiMER (Hierarchical Memory with Elastic Retrieval) through two stages. In Stage I, HiMER constructs a four-level hierarchical multimodal memory spanning token, block, page, and document representations to preserve structural relationships while reducing unnecessary context expansion. In Stage II, HiMER dynamically adapts retrieval granularity according to the query and evolving reasoning state while incorporating a grounding-consistency objective to improve evidence-aligned generation. Experiments on MMLongBench-Doc and SlideVQA demonstrate consistent improvements over SOTA baselines, improving accuracy by up to 6.1%, increasing claim-level groundedness from 64.2% to 81.7%, and reducing retrieved-token cost by up to 71.5%.

Hyperspectral image classification (HSIC) remains challenging due to high spectral dimensionality, limited labeled samples, class imbalance, and strong spectral ambiguity. Although Graph Convolutional Networks (GCNs) effectively exploit contextual relationships between pixels, their performance often degrades under scarce supervision. To address these limitations, we propose a semi-supervised graph-based framework that integrates centroid-driven contrastive learning into a GCN architecture. The proposed method combines a Centroid Metric Loss (CML) and a Triplet-Centroid Loss (TCL) to encourage intra-class compactness and inter-class separability in the embedding space. Class centroids are progressively refined through an Exponential Moving Average (EMA) strategy and leveraged within a confidence-based pseudo-labeling mechanism to exploit unlabeled samples. In addition, a curriculum-inspired weighting scheme gradually strengthens the influence of contrastive supervision during training. Experimental results on three benchmark hyperspectral datasets demonstrate that the proposed framework consistently improves classification performance and robustness compared with several state-of-the-art graph-based and contrastive learning approaches.

Encrypted traffic classification has become increasingly important for network management and security with the widespread adoption of encryption protocols on the Internet. However, existing methods typically rely on single-granularity feature representations and heavily depend on large labeled datasets, leading to insufficient representation learning and degraded performance under low-label scenarios. In this paper, we propose a Semi-supervised Contrastive leArning method for Limited-label Encrypted traffic classification (SCALE). SCALE employs a multi-granularity feature extraction module that progressively extracts discriminative features from the byte level to the flow level to form comprehensive representations. Moreover, SCALE integrates supervised contrastive learning with semi-supervised learning, which enriches the representation space of limited labeled data and exploits abundant unlabeled data, thereby reducing label dependency. Experiments on three public datasets demonstrate that SCALE consistently outperforms state-of-the-art methods across various ratios of labeled data. Even with only 10% labeled data, SCALE achieves an average accuracy of 95.64%, exceeding the strongest baseline on each dataset by 18.05% on average.

Toxic memes often convey offensive, insulting, discriminatory, or threatening meanings through the interaction between visual content and overlaid text. Recent studies have introduced external knowledge generated by large language models or large multimodal models to improve meme understanding. However, generated knowledge is not always reliable; it may contain redundant descriptions, missing cues, hallucinated content, or mismatched reasoning, thereby introducing noise while providing useful knowledge. In this paper, we propose a Multi-View Knowledge-Enhanced Confidence Gating Multimodal Network (MVK-CGMN) for toxic meme detection. The proposed framework constructs text-view, image-view, and joint-view knowledge, and assigns them to different learning paths. Text-view and image-view knowledge guide the unimodal branches through Knowledge Alignment, while joint-view knowledge is selectively injected into the fusion branch by Dual-view MAF (Multi-view Adaptive Fusion). Moreover, a causal-inspired Deconfounder performs representation-level debiasing in the fusion branch, and a Confidence Gate dynamically integrates predictions from text, image, and fusion branches. Experiments on ToxiCN-MM and MET-Meme-CN show the effectiveness of the proposed framework.

Multi-behavior Sequential Recommendation (MBSR) has achieved remarkable success by modeling the sequential dynamics and heterogeneous characteristics of diverse user behaviors to capture users' multifaceted interests. However, existing MBSR methods generally assume that target behavior decisions are uniformly influenced by all historical behavior types, overlooking the personalized and differentiated decision dependencies among heterogeneous behaviors. Consequently, they tend to learn homogenized behavioral representations and fail to accurately capture the varying importance of different behaviors for target decision making. To address this issue, we propose a novel \textbf{D}ecision-aware \textbf{D}ata \textbf{R}ecalibration (\textbf{DDR}) framework that dynamically recalibrates heterogeneous behavioral sequences according to users' personalized decision patterns. Specifically, DDR first partitions the original heterogeneous interaction sequence into behavior-specific subsequences to enable independent decision modeling. It then jointly captures shared cross-behavior dependencies and behavior-specific decision patterns, while employing a target behavior-guided gating mechanism to estimate the personalized contribution of each behavioral subsequence. Based on these learned dependencies, DDR generates recalibrated weights to reconstruct the original interaction sequence, allowing downstream MBSR models to focus on behavior interactions that are more informative for target behavior prediction. Extensive experiments on three real-world benchmark datasets demonstrate that DDR consistently improves the performance of a wide range of state-of-the-art MBSR models, validating both its effectiveness and strong generalizability.

Early rumor detection aims to identify event veracity before social media propagation structures are fully observed. Existing graph neural approaches often pursue stronger graph encoders, implicitly assuming that early propagation graphs already contain stable hierarchical topology. This paper revisits 2,308 propagation trees from Twitter15 and Twitter16 and observes a shallow-wide regime, where over 90\% of reply nodes are direct first-hop responses to the source post and over 97\% lie within two hops. Motivated by this observation, we propose CA-MVP, a classifier-aware multi-view propagation adaptation framework. CA-MVP organizes source semantics, top-down diffusion, bottom-up feedback, and bidirectional aggregation as lightweight propagation views, and couples multi-view representations with strong linear decision boundaries. This view-boundary coupling is designed to mitigate representation homogenization in shallow graphs and boundary collapse in sparse correlated feature spaces. Under a leakage-free early-window protocol, we conduct 50 paired evaluations across two datasets, five observation windows, and five random seeds. CA-MVP significantly outperforms an internal strong baseline and achieves statistically indistinguishable overall performance compared with reproduced Bi-GCN, while maintaining lower training and inference costs. Further diagnostic analyses reveal transition-window selection risks and suggest that near-source replies may contain useful social semantic signals.

Large Language Model (LLM)-based multi-agent systems have shown promise in quantitative finance, where decisions must be made over heterogeneous, high-velocity Web data streams such as financial news, filings, and market feeds, yet they frequently suffer from high inference latency, non-differentiable decision pipelines, and limited integration of risk-management constraints. We recast multi-agent financial decision-making, typically handled as a sequence of LLM deliberations, as a differentiable and risk-constrained control problem executed over a Web-scale parallel streaming substrate. The resulting framework, Distributed Streaming QuantAgents (DS-Quants), couples three mechanisms: a distributed streaming architecture in which specialized Agent-Processors communicate asynchronously over a high-throughput message bus, replacing sequential meetings with parallel execution; a Memory-Augmented Markov Decision Process (MA-MDP) whose differentiable memory retrieval and fusion mechanism enables end-to-end policy optimization in place of black-box LLM prompting, removing LLMs from the decision loop; and a risk-constrained objective that folds transaction costs and a Conditional Value-at-Risk (CVaR) constraint into learning through Lagrangian relaxation. On NASDAQ-100 constituents (2021–2023), DS-Quants attains the strongest risk-adjusted performance among the evaluated methods, with an annualized return of 37.5% and a Sharpe ratio of 2.34 at the lowest maximum drawdown (10.6%), improving modestly but consistently over the QuantAgents baseline. Ablation studies confirm the contribution of the specialized Agent-Processors and the differentiable memory, while the LLM-free streaming design reduces inter-agent decision latency by more than an order of magnitude relative to the LLM-based sequential baseline. DS-Quants offers a step toward a differentiable, LLM-free alternative to LLM-based multi-agent trading, recasting it as risk-aware control.

Multi-agent systems deployed with Large Language Models (LLMs) have significantly advanced collaborative reasoning capabilities. However, existing multi-agent aggregation mechanisms predominantly operate on the answer-level, which discards intermediate reasoning processes and shows vulnerability to malicious or hallucinated Byzantine disruptions. To bridge these gaps, we propose BMTC, a novel Blockchain-based Multi-agent Trajectory Consensus framework designed to secure and align complex, long-term reasoning in decentralized environments. The proposed trajectory consensus distinctly shifts the aggregation baseline from isolated final answers to complete reasoning trajectories. By evaluating the semantic similarity of these trajectories alongside dynamic agent reputations, this mechanism successfully aggregates execution results while accurately identifying and isolating potential malicious agents. Furthermore, we leverage a decentralized blockchain infrastructure to establish an immutable trajectory auditing layer, thereby addressing potential tampering risks in practical deployments. Extensive evaluations on mathematical reasoning benchmarks demonstrate that BMTC achieves superior consensus accuracy and exceptional adversarial resilience.

Personalized course recommendation on Massive Open Online Courses (MOOCs) platforms aims to help learners discover courses that match their learning preferences from a vast pool of resources. Deep learning-based methods rely on sparse learner-course interaction data and have long suffered from representation degradation; recent research has attempted to alleviate this issue by introducing Large Language Models (LLMs), yet existing methods still face limitations in two key aspects: external knowledge introduction and collaborative representation learning. Specifically, the former encodes course semantics independently from text and lacks explicit modeling of structural relations among courses; the latter fails to effectively improve collaborative representations, leaving the representation degradation problem unresolved. To this end, with the goal of improving the quality of learner and course representations in sparse scenarios, we propose SEPAL, an LLM-enhanced semantic-collaborative representation dual-level alignment framework for course recommendation. The framework designs a structure-enhanced semantic modeling module that injects course hypergraph structural priors into LLM semantic representations via hypergraph convolution, thereby providing structure-aware course representations that fuse high-order associations for subsequent prototype-level alignment. It further introduces an instance-level and prototype-level dual-level alignment mechanism: instance-level alignment aligns the local consistency between semantic and collaborative representations at the individual course and learner level, while prototype-level alignment aligns the global cluster structure of the two at the group level, thus improving representation robustness in sparse scenarios and mitigating the representation degradation problem. Experiments on MOOCCube demonstrate that SEPAL outperforms existing baselines on HR@$K$ and NDCG@$K$ metrics.

Missing values are prevalent in time series data and sub- stantially impair the reliability of downstream tasks in real-world appli- cations. Diffusion-based models have recently shown promising results for time series imputation; however, in conditional time series imputa- tion, missing masks and non-stationary temporal dynamics may induce effective perturbations that are mismatched with the Gaussian prior as- sumed by standard diffusion objectives. Furthermore, existing methods struggle to model evolving spatial correlations and multi-scale tempo- ral dependencies effectively. In this paper, we propose NSCDM, a noise self-calibrated diffusion model for time series imputation. Specifically, NSCDM introduces a noise distribution calibration mechanism to align the effective perturbation distribution with the Gaussian diffusion prior. In addition, NSCDM employs a dual-branch architecture that jointly models temporal and frequency representations. By fusing these com- plementary representations with an attention mechanism, it can capture evolving spatial correlations and multi-scale temporal patterns in the time series. Experimental results on multiple real-world datasets demon- strate that NSCDM consistently outperforms state-of-the-art imputation methods in terms of accuracy and robustness.

Vision-Language Models (VLMs) expand the attack surface of safety-aligned systems by coupling visual perception with text generation. Existing multimodal jailbreak attacks primarily rely on crafted visual content, adversarial perturbations, or image-specific attack strategies, leaving the potential of reusable visual anchors in benign natural images largely unexplored. To address the problem, we introduce \textbf{MemJack}, a memory-augmented multi-agent framework for automated VLM red-teaming with natural images. MemJack combines visual-anchor discovery, visual-semantic camouflage, response evaluation, reflection-guided repair, and dynamic replanning within a closed-loop attack pipeline. A lightweight Iterative Nullspace Projection (INLP) filter screens candidate prompts for premature refusal signals, while a persistent Multimodal Experience Memory retrieves and updates reusable anchor--goal--strategy mappings across images. Beyond attack generation, MemJack automatically transforms public images into attack anchors, enables the construction of \textbf{MemJack-Bench}, a dataset of over 113,000 interactive multimodal jailbreak trajectories for safety evaluation and defensive alignment. Extensive empirical evaluations across full, unmodified COCO val2017 images demonstrate that MemJack achieves a 71.48\% attack success rate (ASR) against Qwen3-VL-Plus, scaling to 90\% under extended budgets. Compared with representative multimodal jailbreak baselines under the same natural-image evaluation setting, MemJack achieves the highest ASR while requiring fewer mean rounds to success, demonstrating superior jailbreak effectiveness on VLMs. These results demonstrate that benign natural images can act as transferable jailbreak anchors and reveal substantial vulnerabilities in current safety-aligned VLMs.

Spiking Graph Networks (SGNs) represent a biologically inspired paradigm that combines the energy-efficient properties of Spiking Neural Networks (SNNs) with the relational modeling capabilities of Graph Neural Networks (GNNs). Despite their potential, existing SGNs are largely confined to static settings with real-valued or binary inputs, lacking the capacity to model continuous and dynamic graph evolution. Moreover, the discrete nature of spike-based representations introduces significant information loss, limiting predictive performance. In this work, we identify a critical gap in the literature: the absence of a unified, theoretically grounded framework for energy-efficient dynamic graph learning that leverages spiking dynamics. To this end, we propose Continuous Spiking Graph ODE Networks (CSGO), a novel framework that embeds SGNs within the continuous-time dynamics of Graph Neural Ordinary Differential Equations (Graph ODEs). CSGO bridges the temporal spike-latency in SNNs with the continuous evolution of graph structures via ODEs, enabling both dynamic modeling and low power consumption. To address the intrinsic information loss in spiking representations, we further derive a second-order spiking mechanism and formulate the second-order CSGO, enhancing representational capacity. We provide theoretical analysis demonstrating the gradient stability of our framework, and validate its effectiveness through extensive experiments on benchmark datasets. CSGO achieves state-of-the-art performance while significantly reducing energy consumption, offering a principled and scalable solution for neuromorphic dynamic graph learning.

Reference-based diffusion style transfer is often limited by unstable coordination between content structure and style evidence during inference. A single reference image may entangle semantic cues with multiscale texture patterns, leading to latent mismatches at the sampling origin, unstable query-granular style routing during attention, and semantic drift across denoising timesteps. We propose StyleVAULT, a training-free framework built on a frozen latent diffusion model to coordinate these three sources of instability during inference. HSDC calibrates content and style latents at timestep $T$ through quantile-guided distribution transfer, channel-wise affine-statistic alignment, and post-affine quantile refinement. UQGC estimates query-granular routing uncertainty from branch dominance, style-response ambiguity, top-response margin, and structural sensitivity, and uses it to regulate query fusion, Top-$k$ style competition, residual style preservation, and output-side modulation in selected self-attention layers. SERM further applies CLIP-guided latent rectification with fixed text anchors after denoising updates. Across diverse style categories, our method achieves the best ArtFID, FID, CFSD, Style Loss, and Aesthetic Score among the evaluated methods. It maintains the best ArtFID ranking at the tested resolutions.

Ensuring robust safety alignment while preserving utility is critical for the reliable deployment of Large Language Models (LLMs). However, current techniques fundamentally suffer from intertwined deficiencies: insufficient robustness against malicious attacks, frequent refusal of benign queries, degradation in generated text quality and general task performance, the former two reflecting tensions in \textbf{robust safety} and the latter constituting \textbf{utility impairment}. We attribute these limitations to the coarse-grained layer-wise interventions in existing methods. To resolve this, we propose \textit{NeuronTune}, a fine-grained framework that \textbf{pinpoints and modulates sparse neurons} to achieve simultaneous safety-utility optimization. Our approach first pinpoints safety-critical and utility-preserving neurons across all layers via attack-aware attribution, then adapts meta-learning to adaptively modulate their activations. Crucially, the intervention scope of NeuronTune is dynamically controlled via neuron inclusion thresholds, providing a flexible mechanism to prioritize either security-critical or utility-priority requirements. Extensive experimental results demonstrate that our method outperforms existing state-of-the-art technologies, achieving superior model safety while maintaining excellent utility.

In recent years, online reviews have become a valuable data source for organizations to improve customer satisfaction. Aspect-Based Sentiment Classification (ABSC) helps to extract meaningful information out of these vast amounts of textual data by identifying the sentiments expressed toward certain aspects in reviews. This process can be automated through the use of Large Language Models (LLMs). Despite their capabilities, LLMs may lack domain-specific knowledge needed to interpret information for ABSC. In this research, we propose an approach based on knowledge fine-tuning LLMs using domain ontologies for ABSC. Six LLMs are evaluated on SemEval datasets in a zero-shot and a three-shot approach. The results show that fine-tuning LLMs using verbalized domain ontologies improves accuracy and weighted F1 scores in the zero-shot approach in almost all instances. In the three-shot setup, larger models tend to benefit more from fine-tuning than smaller models.

Recent studies have explored the use of large language models (LLMs) to extract predictive signals from financial news for stock forecasting. However, two gaps remain in current solutions: retrieved evidence must be relevant, timely, and relationally grounded, while the mapping from correctly interpreted evidence to stock returns must remain effective as market conditions change. In this paper, we bridge the gap by proposing STEPS MATA, a temporal knowledge system for LLM-based stock movement classification. We first introduce STEPS (Semantic Temporal Evidence Path Sampling), a graph sampling method that prioritizes news based on semantic and temporal relevance within a temporal knowledge graph. During inference, we propose MATA (Multi-Aspect Temporal Adaptation) to decouple relatively stable semantic assessment from the non-stationary mapping between financial signals and subsequent returns. We evaluate the framework on two real-world stock sets, NASDAQ100 and CSI100 under two distinct financial news environments. The framework connects timely relational evidence retrieval with structured multi aspect financial reasoning and market-regime adaptation.

Reliable prediction intervals are essential for deploying time-series forecasting systems in high-stakes domains, yet uncertainty estimates from frozen large language models (LLMs) often deviate from nominal confidence levels. This paper studies interval calibration for multi-step forecasting with frozen LLMs and proposes CP-LLM, a fine-tuning-free two-stage post-hoc framework. In Stage 1, base intervals are constructed from sampled LLM forecasts and augmented with three complementary uncertainty signals: sampling dispersion, temperature sensitivity, and serialization perturbation sensitivity. In Stage 2, base intervals are calibrated on a held-out calibration segment using either Conformalized Quantile Regression (CQR) for relatively stable regimes or Adaptive Conformal Inference (ACI) under temporal distribution shift. Experiments on five public univariate datasets with four API-accessed frozen LLMs show that CP-LLM substantially reduces coverage bias while maintaining competitive interval sharpness. Additional comparisons with traditional forecasting methods, together with ablation and parameter-sensitivity analyses, provide further evidence for the effectiveness of the proposed framework.

Cardinality estimation is important for a query optimizer to generate high quality query plans in database systems. In recent, pre-trained cardinality estimators, a type of ML-based models trained on diverse datasets sourced from multiple databases, have achieved state-of-the-art accuracy in zero-shot scenarios on unseen workloads. However, the complex relationship between queries and their cardinalities makes it difficult to simply measure the impact of datasets on pre-trained cardinality estimators. In this paper, we propose a framework named Cardinality Estimation Dataset Reweighting (CEDaRe), which includes two phases: CE reweighting and CR sampling, and is designed to improve the performance of pre-trained cardinality estimators by refining datasets. In CE reweighting phase, a kind of group distributionally robust optimization is proposed to optimize the composition of pre-training datasets. Then in CR sampling phase, given the high cost of acquiring training data, an improved stratified sampling method is proposed to sample data according to the computed reweights, which determines the final dataset for pre-trained cardinality estimators. Experimental results on the IMDB, ErgastF1 and Genome datasets show that CEDaRe improve the 99\% Q-error by 1.58x and the End-to-End time by 1.09x at most over the baseline model Price trained using its default datasets weights. Extensive experiments demonstrate that our optimization is both generalizable and robust.

Point-of-Interest (POI) recommendation aims to predict users’ next locations based on their check-in histories, where spatio-temporal information plays a crucial role in shaping user mobility behaviors. Prior works typically treat spatio-temporal differences as auxiliary numerical features or adopt simple linear embeddings, and often entangle spatio-temporal signals with other features, which limits their performance and ability to capture semantic meaning and logical relationships, such as temporal precedence and spatial proximity. In parallel, no existing work employs neural-symbolic first-order logic reasoning to define and implement POI recommendation. To this end, we propose STLRec, a Spatio-Temporal Logical reasoning framework for POI Recommendation with multi-view signals from user check-ins and spatio-temporal contexts. Specifically, it first introduces a feature reasoning module, which captures user-specific preferences by combining semantic attributes and check-in patterns. Then, a spatial reasoning module models geographical relationships between POIs, enabling range-aware spatial inference. Meanwhile, a temporal reasoning module captures temporal dependencies and sequential behavior. Experiments on real-world datasets show that STLRec achieves superior performance compared to existing methods, highlighting the effectiveness of our framework.

Retrieval-Augmented Generation (RAG) has shown strong potential for educational recommendation, yet standard pipelines primarily rely on textual similarity and do not explicitly model the pedagogical relations that make a recommendation useful, such as prerequisite-like progression, difficulty alignment, and topic continuity. We propose a Knowledge Graph-augmented RAG framework that combines dense retrieval with graph-guided evidence expansion, using a lightweight educational knowledge graph built from course metadata, including skills, categories, and difficulty levels. Beyond ranking performance, we introduce relational grounding as an evaluation perspective for assessing whether generated recommendations are supported by explicit pedagogical relations, not only by retrieved text. We evaluate the framework against Vanilla RAG and category-aware RAG across three open-weight generators on a conditioned generation benchmark of 9,000 queries for which the held-out target is available in the retrieved candidate pool, thereby isolating generation and reranking from retrieval coverage. On this subset, KG-RAG consistently improves recommendation quality, reaching up to 94.0% HR@5 and 82.3% NDCG@5, with the largest gains on progression-oriented queries. A structured grounding audit further shows that KG-RAG achieves high relational support, with joint grounding between 92% and 96%, while keeping unsupported claims rare.

Predictive process monitoring (PPM) aims to forecast the future states of ongoing business processes. Although deep learning techniques have been widely used in this domain, they often fail to fully capture the rich semantics implicit in process data. The powerful semantic comprehension capabilities inherent in Large Language Models (LLMs) open up new opportunities for PPM. However, existing studies are often tightly coupled with specific event logs, prediction tasks, and customized reasoning paradigms, which limits the potential of LLMs for generalization and knowledge transfer. In this paper, we propose LLM4PPM, a novel framework that learns transferable process semantics. First, using semantic stories generated from event logs across diverse domains, we perform self-supervised continuous pre-training, enabling the LLM to capture stable and intrinsic process evolution logic from large-scale corpora. Subsequently, we design a multi-task instruction fine-tuning paradigm that consolidates three core PPM tasks, namely next activity prediction, remaining time prediction, and final outcome prediction, into a unified generative LLM. Extensive experimental results demonstrate the effectiveness of the proposed framework.

Component-level audio deepfake detection requires reasoning about both foreground speech and background environmental sounds, which may be independently manipulated. Although the task is commonly evaluated as five-class classification, the target labels are induced by a compositional structure involving the original/mixed state, speech spoofness, and environmental spoofness. In this work, we show that a key challenge under domain shift is not merely the absence of useful acoustic evidence, but the mismatch between frozen expert evidence and the target-domain decision boundary. To address this issue, we propose CE-Router, a Compositional Evidence Router for component-level audio deepfake detection. CE-Router reuses frozen expert detectors as evidence sources, constructs a structured evidence vector from global class predictions, original/mixed cues, confidence statistics, cross-expert agreement, and speech/environment spoofness scores, and learns a lightweight evidence-to-decision router without updating any acoustic backbone. Experiments on CompSpoofV2 show that source-domain routing and naive expert fusion fail to transfer well, while label-efficient routing substantially recovers target-domain performance. With only 200 and 500 labeled samples per class, CE-Router achieves 0.8254 and 0.8397 Macro-F1, recovering 75.4% and 84.3% of the gap between source-domain routing and the full-label upper bound. Evidence-source ablations, calibration analysis, and routing behavior diagnostics further confirm the complementary roles of global, mixed-state, and component-spoofness evidence.

Forced-choice (FC) tests are widely used in online noncognitive assessment to reduce acquiescence and social-desirability biases. To further improve testing efficiency, Forced-Choice Computerized Adaptive Testing (FC-CAT) adaptively selects item blocks based on respondents' estimated traits instead of administering a fixed test. Existing FC-CAT methods still face two limitations. First, many approaches rely on manually designed diagnostic models and handcrafted information criteria, restricting the learning of adaptive selection strategies directly from response data and potentially leading to suboptimal long-term decisions. Second, existing adaptive testing methods are often not tailored to the ranking-based and comparative nature of FC responses, limiting item-selection optimization under FC-specific constraints. We present Forced-Choice Neural Adaptive Testing (FCNAT), a reinforcement learning framework that learns a policy for adaptive FC block selection. The agent encodes each respondent's cumulative ranking history as a state vector, applies action masking over already-asked blocks, and is optimized with a ranking-oriented reward emphasizing informative comparisons and coverage. FCNAT is agnostic to the underlying FC diagnostic model and can be integrated with different FC diagnostic models. Experiments on real-world datasets demonstrate that FCNAT consistently improves diagnostic performance while enhancing testing efficiency through adaptive block selection.

Multimodal sentiment analysis aims to predict emotional polarity by integrating text, audio, and visual modalities. However, multimodal data in real-world scenarios are often incomplete, giving rise to the task of Incomplete-modality Multimodal Sentiment Analysis (IM-MSA). Existing approaches typically focus on either reconstructing missing content or modeling cross-modal emotional correlations, yet fail to jointly model both, leading to semantically inconsistent or affectively misaligned representations. To address this issue, we propose HR-MoE, a Hyper-Reconstruction Mixture-of-Experts framework that reconstructs missing content and models cross-modal emotional correlations simultaneously. HR-MoE employs an implicit modality reconstruction module to restore missing content, along with an adaptive expert mixing and fusion mechanism to model emotional interactions. Besides, a two-stage training strategy is introduced to guarantee the efficacy of HR-MoE, which first aligns modality experts with contrastive learning, then jointly optimizes the reconstruction and fusion processes. Our approach ensures semantically complete and emotionally coherent representations, effectively over coming the fundamental limitations of prior methods. Extensive experiments on 3 widely used datasets demonstrate that HR-MoE outperforms state-of-the-art baselines across various missing-modality scenarios.

Cross-domain recommendation (CDR) aims to alleviate the cold-start problem by transferring user preferences from a source domain to a target domain. Among existing methods, mapping-based approaches have become a popular way to realize cross-domain preference transfer. However, existing mapping-based methods mainly fall into two categories with inherent limitations. Discriminative regression approaches learn rigid point-wise mappings that struggle to capture complex distributional shifts across domains. Meanwhile, diffusion-based generative approaches rely on noise-to-data generation, resulting in high inference costs and an indirect transfer of source-domain preferences. To address these limitations, we propose Flow Cross-Domain Recommendation (FlowCDR), in which we introduce flow matching to model a direct preference evolution process for cross-domain recommendation. Specifically, FlowCDR leverages paired source and target embeddings of overlapping users to learn a continuous velocity field that directly evolves source-domain preferences toward target-domain representations. This design enables deterministic transfer and efficient inference with only a few ODE integration steps. Extensive experiments on three Amazon benchmarks demonstrate that FlowCDR consistently outperforms representative mapping-based and diffusion-based CDR methods in recommendation accuracy, while achieving a more favorable trade-off between accuracy and efficiency. Our code is available at https://github.com/00starry00/FlowCDR.

Multi-behavior recommendation improves target-behavior prediction by exploiting heterogeneous user feedback, while multi-modal item information provides complementary semantic clues for understanding user intent. However, effectively integrating multi-behavior and multi-modal information remains challenging. (i)Auxiliary-target modal intent mismatch: auxiliary and target behaviors may rely on different modal cues, so directly transferring auxiliary behavior signals can introduce mismatched preferences. (ii) Behavior-context-dependent modal reliability: textual and visual priors are not consistently reliable across user-item-behavior contexts, and fixed fusion may amplify noisy or target-irrelevant semantics. To address these challenges, we propose \textbf{Reliability-Guided Stable Modal Transfer (RGSMT)} for multi-modal multi-behavior recommendation. RGSMT constructs target-aware behavior contexts from behavior-specific histories, estimates sample-wise behavior-modal reliability for textual and image priors, and decomposes modal priors into stable and noisy components. It then selectively injects reliability-weighted stable semantics into the target item representation, enabling reliable auxiliary-to-target semantic transfer for target-behavior prediction. Experiments on two real-world datasets demonstrate that RGSMT consistently outperforms state-of-the-art recommendation baselines. Our code is available at https://anonymous.4open.science/r/TOSMT-53C1/.

Approximate nearest neighbor search (ANNS) is critical for large-scale vector retrieval. Product Quantization (PQ) achieves high memory efficiency by decomposing vectors into subspaces, but its performance heavily depends on how dimensions are partitioned. Existing strategies often rely on fixed partitions, global rotations, or variance-based statistics, which may fail to capture the geometric structure of projected data and lead to suboptimal subspace assignments. In this paper, we propose Adaptive Product Quantization (APQ), a geometry-aware framework that learns subspaces using covering complexity. APQ initially treats each dimension as a separate one-dimensional subspace and iteratively merges the pair that incurs the smallest increase in covering complexity, thereby forming non-uniform subspaces that remain relatively easy to quantize. We then design an efficient estimator and show that bounded estimation error preserves well-separated merge decisions. Based on the learned partition, APQ assigns variable codebook sizes according to each subspace's estimated covering complexity and builds a compact index. During query processing, APQ employs a pruning technique to accelerate search. Experiments on real-world datasets show that APQ improves the recall-throughput trade-off over quantization baselines, reduces indexing time, and achieves substantial pruning with small recall loss.

Multivariate time series forecasting (MTSF) relies on expressive local representations and reliable cross-variable dependency modeling. Although recent patch-based methods improve fine-grained dependency learning through dynamic spatial-temporal filtering, they often learn patch representations mainly in the time domain and overlook temporal-distance priors during dependency filtering. To address these issues, we propose FreqTimeNet, a frequency-enhanced and time-aware framework for patch-level MTSF. It first segments series into patches and builds temporal patch embeddings through linear projection. For each patch, rFFT is employed to produce frequency representations with optional DC-component suppression. Based on these embeddings, a time-aware graph filtering encoder incorporates temporal distance into patch-level graph learning through learnable decay weights, refining dependency filtering across time and variables with an explicit temporal prior. Finally, an MLP maps learned representations to future horizons, with reversible instance normalization used to restore forecasts to the original scale. Experiments on 9 real-world datasets demonstrate that FreqTimeNet achieves competitive performance against state-of-the-art baselines, with relative reductions of up to 3.47\% in MSE and 3.98\% in MAE over the strongest baseline in individual settings.

Domain-generalized fact verification aims to evaluate the veracity of out-of-domain claims exhibiting distribution gaps relative to the training data. Existing methods, relying on data augmentation or representation alignment, struggle when target-domain data is unavailable during training. Some approaches propose learning domain-invariant features from source domains to generalize effectivey to unseen target domain. However, coarse-grained representation optimization fails to capture crucial domain-invariant information for claim verification, and overlooks the unique characteristics of source domains that can provide valuable guidance for generalization. Motivated by the issues, we propose Multi-granularity Inference Unit Feature Disentanglement framework (MIUFD) for domain-generalized fact verification. Our framework comprises four components: (1) an inference unit decoupling mechanism modeling logical relationships at multiple granularities; (2) a contrastive-enhanced module for domain-invariant feature learning via joint domain-label constrained alignment; (3) an expert network–based module for adaptive domain-specific knowledge extraction; and (4) a heterogeneous coupling layer fusing multi-channel features for the final verdict prediction. Extensive experiments on four public datasets demonstrate the superiority of our proposal in out-of-domain claim verification. Notably, MIUFD achieves average improvements of 7.82% and 6.33% over the state-of-the-art baseline in terms of F1-score and accuracy, respectively.

Efficient 3D visual grounding aims to localize the target object in a 3D scene according to a free-form language description. Recent sparse convolutional grounding frameworks improve single-stage efficiency by pruning active voxels before expensive cross-modal interaction. However, when processing sparse, noisy, or incomplete point clouds, existing efficient grounding models may still overlook weak but structurally important voxels, produce locally inconsistent features in recovered regions, and treat visual channels with different reliability uniformly during voxel-text matching. To address these challenges, we propose TIP-TSP3D, a sparse-convolutional framework that improves the reliability of text-guided voxel pruning and subsequent vision-language alignment. Specifically, we design a Discrete Morse Theory-guided Topological Pruning strategy to identify critical voxels and bridge-like structures and preserve them during text-guided pruning. We further develop a Harmonic Field Continuity Completion module to perform harmonic graph diffusion on the sparse voxel graph, recovering locally consistent features around retained and completed regions. Finally, an Adaptive Information-Geometric Attention mechanism incorporates a lightweight channel reliability estimate into cross-modal attention, reducing the influence of unstable visual channels during language-conditioned alignment. Built upon the TSP3D baseline, TIP-TSP3D improves ScanRefer [email protected] from 46.71% to 50.25%, corresponding to a 7.58% relative improvement, while maintaining real-time inference speed.

Anonymization enables privacy-preserving data publishing by removing identifiers and generalizing quasi-identifiers before release. In the anonymization process, data are released by different data owners using their own anonymization methods. However, the released anonymized data can still suffer from privacy leakage caused by composition attacks. Existing composition attacks rely on exact value matching. These attacks cannot capture the links when generalized values are expressed differently but have similar meanings. To address this issue, we propose the SemLink framework, the first LLM-based framework for semantic linkage attacks in multi-release anonymized data. We reveal that semantic linkage risk is a practical privacy threat in multi-release anonymized data. SemLink leverages the semantic reasoning capabilities of LLMs to identify privacy leakage arising from such links. SemLink retrieves semantically similar record pairs, constructs and prunes cross-release record chains, uses an LLM to verify semantic consistency, and report high-risk links according to predefined criteria. Experiments on real-world datasets show that SemLink achieves the highest precision among all compared methods and outperforms the baselines by \textbf{8.84$\times$--68.44$\times$}. SemLink effectively detects semantic linkage risks that traditional methods miss.

A/B testing serves as the gold standard for data-driven decision-making across large-scale online industries. To enhance experimental sensitivity and mitigate metric variability, practitioners frequently employ variance reduction techniques such as control variates and regression adjustment. Although numerous variations of these methods have emerged over the past decade, a comprehensive understanding of their theoretical relationships and comparative advantages remains underexplored. This paper provides a rigorous evaluation of their statistical properties and establishes a formal connection between these two frameworks within the context of real-world applications. Beyond the traditional design-based paradigm, we extend our analysis to the model-based framework to identify potential discrepancies in estimation. Our research demonstrates that variance estimators typically considered conservative in design-based settings can become unreliable when applied in model-based contexts. These theoretical findings are supported by extensive simulations and large-scale empirical experiments.

Fairness-aware task assignment in spatial crowdsourcing faces a fundamental tension: existing approaches either rely on global system scans that cannot scale to city-wide operations, or duration-blind heuristics that misrepresent workers with heterogeneous shift lengths. We propose a two-phase localized pruning-and-ranking framework that resolves both challenges at O(k) complexity per event. A grid spatial index first prunes the candidate set to k nearby participants; a strategy-specific fairness instrument then ranks within this local window. We instantiate two strategies: k-Nearest Least-First (k-NLF), targeting task-count equity, and a Composite Dispatcher using an EWMA of worker idle time for duration-aware equity. Evaluated on DiDi GAIA and Gowalla, k-NLF improves worker fairness (JFI) by up to 17.4% over Greedy with zero throughput cost and a 15.7x runtime advantage over LP baselines. The Composite dispatcher achieves the lowest wall-clock runtime of all evaluated strategies (112 s) while maintaining competitive fairness, demonstrating that locality-constrained temporal fairness signals are computationally compatible with real-time dispatch requirements.

Incomplete time series clustering (ITSC) is a fundamental yet challenging problem in the data mining community. Existing two- stage approaches decouple imputation and clustering, often leading to suboptimal performance due to error accumulation, while recent end-to- end methods typically lack theoretical guidance and fail to effectively balance reconstruction quality with clustering objectives. To address these limitations, we propose a unified framework, DIB-ITSC, which in- tegrates diffusion models with the Information Bottleneck (IB) principle for ITSC. Specifically, a bidirectional LSTM with a self-attention mech- anism is employed to learn latent representations from incomplete obser- vations. Meanwhile, the IB objective constrains the latent space to pre- serve cluster-relevant information while discarding redundancy. To mit- igate error propagation and improve robustness, we perform imputation in the latent space using a diffusion model, which progressively refines representations through a denoising process. For clustering, DIB-ITSC incorporates fuzzy clustering into the IB framework to further handle ambiguous boundaries for clustering performance. The entire framework is trained in an end-to-end manner by jointly optimizing imputation and clustering objectives. Extensive experiments on benchmark datasets demonstrate that the proposed method consistently outperforms state- of-the-art baselines, validating its effectiveness for incomplete time series clustering.

Multi-source Web forecasting services fuse local histories, online environmental feeds, socioeconomic panels, event calendars, metadata, and relations among entities. A practical risk in such services is negative relational transfer: relations learned from observed locations may degrade prediction for a new region. In this paper, we study this risk in strict regional induction for urban carbon forecasting, where the target region is excluded from forecaster fitting, train-time graph aggregation, routing, and calibration. We formulate graph use as signed conditional utility relative to an explicit non-relational fallback and propose CauGUARD, a sample-level graph utility routing framework. CauGUARD learns utility from source-region pseudo-targets, ranks candidate graph experts for each sample, and returns the fallback unless the predicted gain exceeds a source-calibrated threshold. Experiments on multi-source urban carbon data show that fixed graph experts can be unreliable in aggregate yet remain useful for selected samples. CauGUARD consistently improves over the fallback, fixed graph experts, and a hard mixture-of-experts router, while safely returning the fallback when graph use is not supported. No-abstention and shuffled-utility controls show that calibrated fallback selection is essential for avoiding negative-utility graph transfer. A secondary chronological check on traffic forecasting provides consistent supporting evidence. These results indicate that reliable Web forecasting requires explicit control of when relational information is transferred, rather than only more expressive graph predictors.

Text-to-SQL over unseen database schemas remains sensitive to unstable lexical correspondence between questions and schemas, particularly under synonym substitution, and must also recover suitable join structures for multi-table queries. In this paper, we propose an LLM-guided method that uses the LLM as a structured prior generator rather than a direct SQL generator. Candidate Schema Recall broadens the initial search space, a query-aware schema prior confirms, filters, and supplements the recalled candidates, and schema-valid join-path segments are converted into explicit inter-table relations for a structured parser. Experiments on Spider and Spider-Syn show a substantially larger improvement over the reimplemented baseline under synonym substitution, with the full model exhibiting larger differences on multi-table queries than on single-table queries. Further analysis reveals a central limitation of relevance-oriented priors: semantically relevant schema items are not necessarily required by the target SQL. The results support using constrained LLM priors as complementary input-side information for controllable structured parsing.

Audit anomaly detection is commonly formulated as a binary classification task, whereas practical auditing requires suspicious vouchers to be prioritized under limited investigation resources. To address this mismatch, we propose Audit-iMAD, a hierarchical framework that retains binary classification for score learning but reframes deployment as risk-oriented prioritization followed by selective knowledge-grounded reasoning. First, Audit-iMAD introduces au-dit-domain knowledge into a MacBERT-based semantic encoder and combines the resulting anomaly score with an independently aggregated graph-risk sig-nal to construct a unified risk ranking. A diversity-aware top-K strategy then reduces redundant candidates and retains representative high-risk vouchers. Second, the framework constructs voucher-specific contexts from applicable audit rules, knowledge-graph evidence, and historical audit cases retrieved through risk-aware retrieval augmentation. Third, a language model performs evidence-grounded reasoning with self-critique and self-consistency, while an uncertainty gate selectively escalates difficult cases to role-diversified multi-agent debate. The evaluation is conducted in a representative specialized audit scenario from a real-world Chinese ERP environment, where the source vouchers have undergone business-oriented filtering, textual cleaning, and ac-counting-subject normalization and collectively form the domain knowledge environment used for screening and reasoning. Within this knowledge-rich op-erational setting, semantic-graph risk fusion improves high-risk screening over the controlled semantic baseline, reaches complete anomaly coverage under a smaller review budget. These results support Audit-iMAD as a practical hier-archical screening framework for large-scale Chinese ERP auditing, while the concentrated business scope and strong template regularity of the current sce-nario limit claims of unrestricted cross-domain generalization.

Routing in Wireless Sensor Networks (WSNs) is continuously challenged by two distinct phenomena: slow-evolving macro-dynamics, such as gradual node depletion and large-scale topological shifts, and rapid micro-dynamics, including transient congestion and temporary channel fading. Addressing both simultaneously within the strict energy budgets of IoT edge devices requires a fundamental paradigm shift. This paper introduces a Hybrid Q-Learning Automata protocol designed explicitly to bridge these operational scales. The protocol integrates a meta-controller that continuously evaluates network volatility, seamlessly shifting between a powerful Q-learning macro-solver for global optimization and an ultra-efficient Learning Automata micro-solver for localized routing. By decoupling long-term gradient mapping from immediate probabilistic forwarding, the protocol maintains high adaptability without the continuous computational overhead of pure value-iteration. Our simulations confirm that this hybrid approach drastically outperforms AODV, RPL, and LEACH in packet delivery and load balancing, offering a Pareto-optimal alternative to energy-intensive Deep Reinforcement Learning.

Recommender systems often suffer from popularity bias, where popular items dominate while niche items receive limited exposure. This bias stems from over-reliance on historical interactions that fail to capture fine-grained user preferences. We propose Expl-Debias, a two-stage framework that mitigates popularity bias via explanation-aware learning. Stage 1 learns base user–item utility from interactions, while Stage 2 incorporates LLM-generated positive and negative explanations through contrastive training, promoting relevant niche items and suppressing popular but mismatched alternatives. Experiments on three real-world datasets demonstrate that Expl-Debias consistently outperforms state-of-the-art LLM-based recommendation and debiasing methods in both accuracy and fairness. Code is available at https://github.com/Sairamvinay/Expl-Pop-Debiaser/.

Reliable in-flight anomaly detection is critical for UAVs in web-enabled service pipelines, but remains challenging because raw telemetry couples control-induced flight dynamics with disturbance-related residual patterns. Existing UAV-specific methods mainly model temporal deviations and may miss global disturbances, while general time--frequency methods can introduce pseudo-structural spectral noise by directly decomposing control-driven telemetry. To address these issues, we propose DyDe-AD, a decomposition-before-detection framework for UAV anomaly detection. DyDe-AD first uses a Mori--Zwanzig-guided DeepKoopman decoupling to separate each telemetry window into a Markov dominant component M and a memory-kernel component P, reducing the frequency interference caused by normal control responses. It then reconstructs M with a Multi-scale Retrieval-Augmented Temporal Reconstructor (MRA-Rec) based on scale-specific prototype retrieval, and reconstructs P with a Morlet-wavelet-based variational autoencoder (Morlet-VAE) for frequency anomaly detection. Experiments on real-world and simulated UAV datasets show that DyDe-AD outperforms 17 representative baselines in F1 score and AUC, while using only 0.47\,M trainable parameters and achieving the highest inference throughput. The source code is available at https://github.com/1447083915/DyDe-AD

Long-term time-series forecasting (LTSF) models encode temporal structure through model-specific choices such as tokenization, spectral operators, and latent layouts. These choices can be effective, but they are difficult to reuse across heterogeneous forecasters. We introduce BAViP-TS, a visual conditioner that leaves the forecasting interface, loss, and native prediction path unchanged. It renders the observed encoder window as an auxiliary image, extracts structural features with a vision encoder, and injects them through bounded, identity-preserving modulation. We evaluate BAViP-TS with five es-tablished LTSF hosts–TimesNet, iTransformer, PatchTST, FEDformer, and TSMixer–on seven benchmarks and four horizons. BAViP-TS obtains lower three-seed mean MSE in all 140 paired settings and lower mean MAE in 138 of 140 settings. Ablations identify pseudo-RGB rendering and FiLM-style modulation as the most consistent choices across hosts. Unlike forecasting pipelines built around large language or vision-language models, BAViP-TS requires no text alignment, retrieval memory, or external semantic input at inference time.

Pass receiver selection is a fundamental task in football analytics, aiming to predict the intended receiver under a given game state. This task is challenging with event-centered freeze-frame observations, a broadcast-like setting that provides only partial and variable player visibility without complete trajectories or stable player identities. The model must therefore reason over anonymous visible candidates, opponent pressure, and recent context under partial observation. To address this setting, we propose a Hierarchical Possession-aware Graph Pointer Network (HPGPN), which formulates pass receiver selection as variable-size candidate prediction over visible teammates. HPGPN jointly models current player interactions, local event context, and possession-level temporal dynamics. It represents the current pass situation with a graph, incorporates fixed event context, and uses dynamic possession history to capture how the attacking sequence evolves. Candidate representations are refined hierarchically by integrating spatial, contextual, and historical evidence, and a glimpse pointer head scores the receiver candidates. Experiments on public football event and freeze-frame data show that HPGPN improves pass receiver selection performance. Ablation studies demonstrate the effectiveness of graph-based interaction modeling, fixed event context, and dual-branch dynamic possession-history modeling.

Automatic detection of web accessibility issues largely relies on verifying code compliance with standards, which is useful for static problems but cannot detect issues that only emerge during user interaction with dynamic content. Prior work has proposed strategies to automate detection of these interaction-based problems without real users, but such strategies remain limited to specific widget types and depend on the presence of semantic attributes. In this work, we explore the feasibility of detecting accessibility issues for screen reader users by simulating keyboard-based user interaction. Our automated approach analyzes changes in web interfaces that occur in response to certain event patterns that we call Accessibility Events (AEvents). To test how simulated interaction can uncover dynamic AEvents, we generated simulations of user interaction on 5 real websites, while applying AEvent detection mechanisms. We compared our results against an automated checker and manual expert inspection. Our approach identified 97 issues, complementing the 28 found through manual inspection and detecting substantially more issues than the automated checker. Because it evaluates the interface in real time and adapts to its changes during interaction, our approach offers a scalable complement to manual testing. These findings lay the groundwork for new accessibility evaluation frameworks capable of identifying dynamic accessibility issues.

A recent trend in addressing encrypted traffic classification is to represent raw traffic by graphs, and design a graph neural network (GNN)-based model as the classifier. However, the existing graph-based traffic representations exhibit the following limitations. First, these methods only utilized the first few packets to construct the graph representation for a traffic flow, leading to a serious loss of information and a worsening of classification accuracy. Second, they adopted unweighted graphs, failing to utilize edge weights to convey traffic information. Finally, the design of the existing traffic graphs lacked semantic interpretability. To solve the above challenges, in this paper, we propose a novel graph-based traffic representation method to efficiently construct several Interpretable Traffic Graph Sketches (ITGSs) for raw traffic. Specifically, we efficiently convert a raw traffic flow to three light-weight graph sketches to capture three kinds of information, i.e., interaction of time slots, interaction of packet size ranges, and interaction between time slots and packet size ranges. The above information is conveyed by the edges as well as the edge weights in the graph sketches, making our traffic representations highly recognizable and interpretable. Based on the above representations, we design a deep neural network (DNN) model to respectively extract hidden features from the different graph sketches, and fuse the extracted features for the final classification. Experiments conducted on five public datasets show the superiority of our method compared to the state-of-the-art methods.

Most existing graph analysis-based encrypted traffic classification methods converted raw traffic to a single graph and then performed classification using graph neural networks (GNNs), which failed to capture the temporal information in the original traffic. To this end, in this paper, we propose an Encrypted Traffic Classification method based on Multi-Scale Graph Sequences (ETC-MSGS) to classify traffic based on not only topological information, but also temporal information. First, the raw traffic flow is converted to two sequences of sub-flows based on different time scales. Next, we design a traffic representation method to convert each sub-flow to a novel Packet Size Interaction Graph (PSIG), obtaining two PSIG sequences. Finally, combining graph neural networks (GNNs) and recurrent neural networks (RNNs), we first extract graph-level hidden features and then capture temporal features from the PSIG sequences. Extensive experiments are conducted on five public datasets, and experimental results show that ETC-MSGS outperforms several baseline methods.

The widespread adoption of TLS-encrypted has reduced the effectiveness of traditional traffic feature-based malicious website detection methods. Recent studies indicate that flow–resource relationships can reveal latent patterns of website loading and resource interactions, but constructing them under TLS~1.3 is extremely challenging. To address this problem, we propose 2AM, an ambiguity-aware multi-modal collaboration framework for robust malicious website detection. 2AM jointly leverages transport-layer encrypted traffic sequence and lightweight application-layer resource metadata observable from the client environment. The framework models website loading behaviors as a structure-aware heterogeneous graph, where flows and resources are represented as heterogeneous nodes while preserving ambiguous flow-resource associations. Based on this representation, 2AM employs heterogeneous graph learning together with node-level feature augmentation and adaptive loss optimization to capture cross-modal interactions. Experiments demonstrate that 2AM achieves superior detection accuracy and robustness in encrypted scenarios. The results highlight the effectiveness of ambiguity-aware structure modeling and multi-modal collaboration for malicious website detection.

Log anomaly detection is vital for large-scale systems, yet existing methods struggle with multi-source heterogeneity. We propose MLogMOE, a framework integrating a Source-Adaptive Hierarchical Router (SAHR) with a Mixture-of-Experts (MoE) architecture. SAHR employs a two-tier, source-aware routing mechanism to automatically direct diverse log data to specialized experts, achieving efficient adaptation. Furthermore, a multi-stage training strategy—incorporating sequential pretraining and a Source-Guided Routing Loss—is designed to ensure expert specialization and router convergence. Experiments on multiple benchmarks demonstrate that MLogMOE achieves state-of-the-art performance and significantly enhances computational efficiency, drastically reducing runtime in large-scale deployments compared to baseline methods.

Diffusion models have been introduced into time series forecasting in recent years, demonstrating strong capabilities in distribution modeling and uncertainty characterization. However, existing methods typically apply diffusion processes directly to the overall forecast residuals, lacking explicit constraints on the temporal structure. To address this, we propose Structure-Aware Seasonal Residual Diffusion (SSRdiff) framework, achieving functional separation between structural components and random disturbances at the modeling level. We perform trend-seasonal decomposition on historical series. The trend component and the predictable seasonal structure are modeled separately by deterministic predictors. Subsequently, diffusion modeling is applied exclusively to seasonal forecast errors, probabilistically characterizing random fluctuations arising from high-frequency disturbances and heteroscedastic noise. Furthermore, we design an uncertainty-interaction conditional mechanism to extract locally time-varying cross-variable dependencies from seasonal forecasts. This constructs structured conditions to constrain the denoising process, mitigating sampling drift and instability caused by unstructured diffusion. Experimental results across eight real-world datasets demonstrate that our method achieves significant improvements in both point and probabilistic forecasts, validating the effectiveness of selective diffusion modeling under structural constraints.

In table data mining, table interpretation (TI) aims to generate accurate and consistent semantic annotations for tabular data to support downstream tasks. TI typically involves two core tasks: column type annotation (CTA) and column relation extraction (CRE). Existing methods struggle to effectively model the inherent row-column structure of tables, which restricts their ability to fully capture structural dependencies. They also often rely on external table metadata for additional semantic information, making them sensitive to metadata quality and less robust across different table scenarios. To address these issues, we propose TISS, a metadata-free TI framework that uses only intrinsic table content. Specifically, TISS constructs a heterogeneous graph to explicitly model the two-dimensional structure of tables and introduces structural-semantic nodes to obtain high-level semantic information. Together with a pre-trained language model (PLM) encoder, the graph enables joint semantic-structural representation learning. Moreover, a lightweight large language model (LLM) initializes the structural-semantic nodes. It generates auxiliary descriptions to enrich their semantic information. We evaluate TISS on two public datasets, WikiTable and SemTab. Experimental results show that TISS consistently outperforms representative existing methods on both tasks, achieving new state-of-the-art performance and validating the effectiveness of the proposed framework.

Predictive auto-scaling uses workload forecasts to adjust resources before traffic changes arrive, yet general time series forecasting benchmarks mainly emphasize average errors and rarely capture abrupt workload changes that determine resource risk. We present TRACE, a real-world enterprise Web QPS benchmark for Internet service elasticity. TRACE releases 200+ anonymized life-service QPS series, standardizes them into comparable forecasting tasks with waveform and business-calendar contexts, and evaluates forecasts from three views: standard forecast error metrics, slope-based abrupt-change-weighted error metrics, and downstream auto-scaling utility. We evaluate 15 recent forecasting methods. Results show that average errors can hide forecasting risks during abrupt traffic changes, and these errors can further translate into downstream under-provisioning and over-provisioning risks.

Pedestrian detection plays a crucial role in computer vision with applications in autonomous driving, surveillance, and public safety. However, real-world dense scenes bring severe challenges, including heavy occlusion, drastic scale variations, and strict real-time requirements. Existing lightweight detectors struggle to balance accuracy and efficiency while often neglecting quality-aware feature modeling and consistency between classification and localization, leading to unstable performance under crowded conditions. To address these issues, we propose DensePed-Lite, a unified framework built on a single principle: under occlusion the network should adapt its behavior to the quality of what it observes rather than assume complete information. This principle is realized at three points where occlusion does the most damage---unreliable confidence scoring (UQE), fragmented spatial coverage (MPSC), and incoherent multi-scale fusion (CTDM)---so that the three mechanisms reinforce one another instead of acting in isolation, all without significantly increasing complexity. Experiments on CityPersons and CrowdHuman validate that DensePed-Lite achieves superior accuracy-efficiency trade-off compared with recent state-of-the-art lightweight methods, making it suitable for real-time deployment in dense pedestrian scenarios.

Learned indexes deliver high efficiency for range queries but offer no support for user-side result verification. This is critical in cloud outsourcing, where a semi-honest server may return tampered or incomplete results. Existing authenticated indexes address verification by aggregating node digests bottom-up, and their efficiency hinges on stable node boundaries and layouts: a single update then invalidates only the few digests along one root-to-leaf path, keeping maintenance local. This assumption breaks under heavy updates. As data distribution drifts, learned indexes retrain routing models and adjust node boundaries, causing structural changes to invalidate authentication metadata across affected nodes and forcing costly reconstruction. To address these issues, we propose VE-index, a workload-aware authenticated dynamic learned index for range queries. We design a layered strategy to decouple query serving from update absorption. Specifically, the Stable Authentication Layer (SAL) generates compact cryptographic proofs for consolidated records, while the Incremental Authentication Buffer (IAB) locally accumulates new writes before merging into SAL. A workload-aware scheduler further switches each data node among StableAuth, MutableAuth, and RebuildAuth according to update activity, buffer occupancy, and model error. For each range query, VE-index generates a verification object that binds SAL range proofs and IAB update evidence to the same signed version root. Experiments on the Uniform, Biased, and Longitudes datasets show that VE-index consistently outperforms existing authenticated indexes in both query and insert throughput, while maintaining a clear performance advantage as the update ratio grows.

Unauthorized use of large language model (LLM) services, often called \emph{shadow AI}, is spreading across enterprise networks, forcing security operators to inventory which LLM services are in use from encrypted network traffic. Yet new services appear continuously, so any practical monitor must recognize known services while detecting traffic from those it has never seen. Existing approaches either fingerprint a fixed, closed set of models or applications, or rely on generic packet features and threshold-based unknown detection; no open-set method models the continuous-time token-generation dynamics that distinguish streaming LLM responses. We present STREAM, an open-set recognition framework that casts the irregular sequence of downstream response packets as a continuous-time process: a time-aware state-space encoder binds its discretization step to the physical inter-token interval, and a dual-branch contrastive metric embedding is paired with tied-covariance Mahalanobis scoring under class-conditional conformal calibration, so that a flow far from every known service is identified as ``an unidentified LLM service is in use'' instead of being absorbed into the closest known label. On real-world traffic spanning up to 11 services in two collection rounds, STREAM keeps AUROC from falling below 91.5 at every openness level and is the best or second-best in all 24 comparison settings against eight open-set baselines drawn from four method families.

AI programming assistants increasingly rely on third-party agent skills for extended capabilities, in which skill capability analysis is crucial as the developers may maliciously declare capabilities beyond their true capabilities. Existing defenses such as code signing and sandboxing fail to simultaneously achieve verifiable permissions and privacy protection. This gap leaves assistants vulnerable to permission escalation and dependency chain poisoning. This paper presents Zkasa, a Zero-knowledge proof-based scheme enabling AI agent skill analysis without compromising data privacy. Zkasa constructs a ZK permission circuit enforcing two constraints with bit-level comparison over the standard capability lattice model. We define two properties for skill analysis/verification: self-consistency and dependency monotonicity, capturing that the skill's declared capability covers all its required capabilities and no skill dependency declares a higher capability than what the parent skill declares, respectively. We formally prove the capability privacy and integrity guarantees of Zkasa. Experimental results demonstrate the lightweight verification cost regardless of the skill size or the dependency count. Moreover, we achieve 100% accuracy in identifying both self-violation and dependency-violation cases.

The underlying web security and trust management depend on the memory safety of the native C/C++ components, like JavaScript engines, and media parsers. Among memory-safety defects, Use-After-Free (UAF) vulnerabilities are especially damaging, since they let attacker-controlled web input escalate into memory corruption or remote code execution. Directed greybox fuzzing (DGF) augments conventional greybox fuzzing with directional guidance toward suspected vulnerable sites, making it a natural fit for such bugs. However, existing directed fuzzers steer almost entirely by target locations. Location-only guidance overlooks two properties that decide whether a UAF actually fires, the temporal order and object identity of the alloc–free–use chain and the lack of any in-loop signal under native execution, where many UAFs stay silent until a sanitizer or replay stage observes them. To address this gap, we present UAFGo, a runtime state-aware directed greybox fuzzer for source-available C/C++ targets. UAFGo instruments specific UAF state-transition points, maintains a pointer-aware finite-state machine during fuzzing and exports compact metadata about the deepest live UAF frontier to the fuzzer, effectively guaranteeing the temporal ordering and object consistency of the UAF lifecycle. When the same tracked object triggers the full UAF chain, UAFGo explicitly aborts execution, turning even a silent target UAF into a fuzz-visible crash. On 13 real-world UAF targets, UAFGo achieves a 3.14X geometric-mean speedup in time to the first effective crash over AFLGo+ASan and remains highly competitive with state-of-the-art fuzzers. In the independent audit, 97.69% of first UAFGo aborts correctly reproduce as ASan-detected memory errors.

Graph fraud detection is a critical task aimed at identifying anomalous nodes within transaction networks, social networks, product review networks, and other graphs, thereby reducing societal losses. To distinguish between the normal and the abnormal nodes, recent studies have analyzed the key role of labeled homogeneous and homogeneous connections. However, due to the necessity for privacy protection, graph data is often divided and stored across various local clients. This division results in the neglect of edges between subgraphs, leading to incomplete neighborhood aggregation. Furthermore, such division also reduces labeled nodes available for training in each client subgraph, posing challenges for current aggregation methods that rely heavily on labeled neighbors. To tackle these issues, we propose a prototype-based federated graph fraud detection method from views of edges and nodes. Given that fraud detection involves only three types of edges (normal-normal, anomalous-anomalous, and normal-anomalous), we construct edge prototypes from known edges to calculate the membership of unknown edges to these three types. Combining this with the node class probabilities obtained from the initial prediction, we identify high-confidence edges and incorporate them into the neighbor aggregation process to improve detection performance. After several epochs of model training on each client's subgraph, we use clustering methods to obtain the features of several node prototypes without leaking privacy. By employing an attention mechanism, nodes in the client's subgraph can aggregate information from neighbors outside the client in the next epoch. Comprehensive experiments conducted on four datasets demonstrate the effectiveness of the proposed method in federated graph fraud detection.

Travel recommendation (TR) plays an increasingly important role in providing diversified travel services to users. This work targets multi-scenario click-through rate (CTR) prediction in TR, aiming to develop a unified model that simultaneously serves heterogeneous scenarios. Existing methods only partially address three intertwined challenges in this setting: the cold-start problem, the difficulty of capturing scenario-specific user interests, and the exposure bias induced by manual operational rules. To jointly tackle these challenges, we propose the Multi-Scenario Debias Unified Network (MSDUN), which integrates three modules within an end-to-end framework: the Interest Projection Network, the Intent Memory Network, and the Intervention Debias Network. Extensive offline experiments on two real-world datasets and an online A/B test demonstrate that MSDUN consistently outperforms state-of-the-art baselines, achieving a 4.15% CTR gain in production deployment.

Next Point-of-Interest (POI) Prediction is a cornerstone of location-based services. However, current models are typically built upon a flawed extrapolation paradigm, predicting a user’s next location based solely on a single, isolated query trajectory. This approach is fundamentally ill-equipped for real-world scenarios where trajectory data are extremely short and limited, leading to poor performance and myopic inference. Inspired by human-like associative reasoning, we propose a paradigm shift and introduce RARE, a novel retrieval-augmented prediction framework that features three components. First, RARE introduces a knowledge base grounded in Optimal Transport (OT) theory to enable principled retrieval. This ensures that the trajectory embedding space is metrically-aware, where distance is a theoretically sound proxy for pattern similarity. Second, RARE employs an efficient diversity-aware retriever to ensure an informative yet concise set of associative trajectories.To reason over this associative knowledge, RARE employs a knowledge-augmented predictor combined with a query-contextual hypergraph and cross-attentive knowledge fusion, which captures high-order correlations between the user’s query and the retrieved trajectories, moving beyond the myopic inference. Extensive experiments on three real-world datasets validate that RARE can outperform twelve state-of-the-art models (including LLM-based models), achieving an average improvement of 5.91% and showing particular strength in challenging sparse-data scenarios.

Multi-view video action recognition has received increasing attention in web video understanding, as videos captured from different viewpoints provide complementary cues for more complete action understanding. However, in real-world web and edge deployment scenarios, only a single-view input is typically available at inference, making it difficult to directly exploit cross-view complementarity. Existing methods either depend on synchronized multi-view inputs, which limits their applicability under single-view inference, or learn view-invariant action representations, which may overlook view-specific discriminative cues. In this work, we propose a novel Cross-view Complementarity-Aware Calibration Network (CCAC-Net) for efficient video action recognition. The proposed network leverages synchronized multi-view videos during training to calibrate single-view action representations, while requiring only a single view during inference. Specifically, a lightweight Temporal Refinement Convolution is introduced to refine temporal action cues within each view while preserving view-specific discriminative information. Then, adaptive view weights are estimated over synchronized training views to aggregate cross-view complementary cues into an action-consistent reference. The difference between each view-level action representation and this reference is further defined as a compensation target, which supervises a Compensation Direction Predictor to learn the compensation direction for calibrating each single-view representation. Experiments on four representative benchmarks demonstrate that the proposed method achieves superior recognition performance under the single-view inference setting. Notably, on NTU RGB+D 120, the proposed method achieves the best Top-1 accuracy while requiring only about 1/13.1 of the parameters and 1/5.0 of the FLOPs of the second-best method, demonstrating its effectiveness and efficiency for resource-constrained video action recognition.

As mobile applications (Apps) have become the primary interface for accessing Web services, a growing number of Web information systems rely on encrypted network communications to deliver personalized content and interactive services. Meanwhile, traffic fingerprinting techniques have increasingly been used to infer application usage and user activities from encrypted traffic, raising significant privacy concerns. These techniques not only extract traffic features from packet-level information commonly used in traditional fingerprinting attacks, but also exploit the characteristic that Apps often follow fixed initialization procedures when launching network connections. They achieve this by observing the temporal relationships among an App's multiple flows to extract features beneficial for fingerprinting attacks, thereby posing a more severe threat to the privacy of mobile application users. We propose Sacred, a novel reinforcement-learning-based application fingerprinting defense that is aware of inter-flow relationships. Sacred leverages a transformer to capture packet-level and flow-level states within a context window and generates low-overhead adversarial perturbations in real time through reinforcement learning. In addition to conventional packet padding, Sacred actively initiates synchronized connections to application-irrelevant websites to disrupt temporal dependencies among multiple flows, enabling effective protection of users' Web-service access privacy against inter-flow-based fingerprinting attacks. Experimental results show that Sacred improves defense success rates by 3.1%~22.9% while reducing overhead by 3.3%~45.2% compared with baseline defenses. An ablation study further demonstrates that the proposed active connection initiation mechanism contributes an additional 10.4% percentage points improvement in defensive effectiveness.

With the widespread adoption of encrypted communication protocols, detecting malicious traffic has become increasingly challenging. Existing encrypted traffic detection systems often rely on large-scale accurately labeled datasets, an assumption that rarely holds in practical deployments. In many real-world scenarios, only limited training data is available and labels are frequently corrupted by automated annotation mechanisms. In this paper, we address the problem of low-quality datasets by proposing MEDAL, a framework employing a four-stage training pipeline. MEDAL utilizes a Dual-Stream Micro-Bi-Mamba backbone to effectively capture micro-level packet fluctuations and macro-level flow dependencies. To mitigate the impact of erroneous annotations, we first construct a noise-robust backbone through self-supervised pre-training using SimMTM and InfoNCE contrastive loss. Subsequently, label noise is systematically identified and corrected via Confident Learning, Area Under the Margin (AUM) dynamics, and KNN consistency verification. To address data scarcity, we synthesize diverse feature vectors within the latent space using a Tabular Denoising Diffusion Probabilistic Model (TabDDPM), which are then mixed with real data via mixup augmentation. We demonstrate the effectiveness of MEDAL through experiments on two public datasets. On the DoHBrw dataset, MEDAL achieves an F1-score of 0.86 ± 0.02, even with only 500 training samples and a noise level of 30%.

Website Fingerprinting (WF) attacks remain a critical threat to user privacy in encrypted networks, with modern deep-learning classifiers achieving over 98% accuracy. Existing traffic shaping defenses face significant challenges in practical deployment: they either require white-box access to the attacker's model or rely on full-flow buffering with uniform obfuscation. To address these challenges, we propose SCARD, a defense that integrates a Conditional Variational Autoencoder (CVAE) with Soft Actor-Critic (SAC) reinforcement learning to enable runtime, burst-level perturbation without white-box access or full-flow visibility. SCARD learns a semantic latent space in which gradient-based saliency analysis identifies critically informative traffic regions, and a SAC agent guided by dual-centroid rewards preferentially injects dummy packets at those positions to shift the latent embedding away from its original class. Extensive evaluation on benchmark datasets demonstrates that SCARD reduces worst-case attack accuracy from over 97% to as low as 7.6% while maintaining bandwidth overhead below 10%, achieving the highest defense efficiency ratio among all evaluated methods.

Multi-user cloud storage auditing is important because outsourced files must be checked with sampled possession assurance, compact public adjudication, and limited disclosure of audit openings. However, existing blockchain-assisted schemes either place costly verification on chain, disclose Merkle openings in calldata, or depend on a semi-trusted auditor to summarize evidence. This paper presents SC-DIV, a smart-contract-driven zero-knowledge integrity auditing system for large-batch multi-user cloud storage. SC-DIV keeps compact commitments and round metadata on chain, while performing response aggregation, leaf-tag verification, Merkle-path checking, and Groth16 proof generation off chain. Its protocol combines contract-state binding, commit–reveal randomness with optional VRF/beacon binding, deterministic no-replacement sampling, coordinator-signed transcripts, SNARK-friendly authentication, and explicit 4 KiB chunk-to-field encoding. The analysis establishes sampled data-possession soundness, transcript accountability, on-chain opening non-publication, and traceability. For 100 users and 20 challenges per user, finalization costs 25.72 × 10⁴ EVM-adjusted gas and submits 384 B instead of about 1.10 MiB of Merkle paths; a normal repeated round costs 46.22 × 10⁴ gas. The 60.8M-constraint circuit completes in 68.4 s with 34.6 GB peak memory, demonstrating predictable off-chain execution with compact, fixed-format finalization.

Missing values are prevalent in multivariate time series and seriously undermine the reliability of data analysis and downstream ap- plications. Recently, diffusion models have shown superior performance to statistical and autoregressive models for time series imputation. How- ever, current diffusion models typically overlook the intrinsic geometric structure of time series in the Symmetric Positive Definite (SPD) mani- fold, which limits their ability to accurately fill missing values. To address this issue, we propose MADiff, a novel manifold-aware diffusion frame- work for time series imputation. Specifically, MADiff achieves diffusion modeling on the SPD manifold by Cholesky decomposition, which pre- serves geometric validity while enabling a stable diffusion process. Fur- thermore, MADiff employs an innovative dual-stream architecture that jointly captures temporal and frequency representations. Through fusing these features with a cross-attention mechanism, MADiff captures both fine-grained temporal dependencies and periodic patterns. Extensive ex- periments conducted on four real-world datasets verify that MADiff sig- nificantly outperforms state-of-the-art methods in imputation accuracy.

Long-term memory systems for LLM agents write dialogue into persistent entries and later retrieve them to answer questions, implicitly assuming that fine-grained meanings remain recoverable throughout the memory pipeline. We refer to this capability as cross-lingual semantic preservation. In cross-lingual settings, however, distinct source-language concepts may be associated with the same coarse target-language representation, making their original distinctions difficult to recover or use. Existing memory benchmarks, being largely English-centric, do not test whether such culturally grounded distinctions remain available at query time. To address this gap, we introduce CSI-Mem, a Chinese benchmark that tests whether memory systems keep apart concepts that are distinct in Chinese but may correspond to the same coarse English label. Evaluating five representative memory systems, we find that all of them perform poorly on several tasks that test fine-grained semantic distinctions. Near-exhaustive retrieval does not guarantee high accuracy, showing that retrieval coverage alone is insufficient to ensure end-to-end cross-lingual semantic preservation.

Multimodal clinical diagnosis integrates patient information, such as longitudinal electronic health records (EHR) and chest X-ray (CXR) images, to support comprehensive disease prediction. Existing multimodal methods encode EHR and CXR into global representations and directly output all disease predictions. However, multimodal models often favor modalities that are easier to learn, leading to modality imbalance. In multimodal diagnosis with EHR and CXR images, repeated historical diagnoses can provide strong predictive shortcuts, causing models to over-rely on EHR-derived priors while underusing localized CXR evidence in the final disease ranking. To address this limitation, we propose VERA-Dx, a visual evidence-guided reranking agent framework for multimodal disease prediction. Specifically, VERA-Dx leverages a general diagnostic agent and a respiratory agent to detect low-consistency respiratory visits, where respiratory evidence may be underused in the general disease ranking. It then directs a visual evidence agent to focus on candidate-specific CXR findings and uses a reranking agent to incorporate this evidence into the global disease ranking. Experiments demonstrate that VERA-Dx consistently improves both respiratory-subset prediction and all-disease diagnosis, highlighting the benefit of explicit visual evidence reasoning for refining multimodal disease ranking.

The Parameter-Efficient Fine-Tuning (PEFT) methods have been extensively researched for large language models in downstream tasks. Among all the existing approaches, the Low-Rank Adaptation (LoRA) has gained popularity for its streamlined design by incorporating low-rank matrices into existing pre-trained models. Though effective, LoRA, as well as its adaptive optimizations, either allocate the same matrix to all the modules or adjust the interior rank of the components based on importance scoring indicators. In this paper, we argue that not all the modules in LLMs are suitable and necessary to be fine-tuned. Enlightened by this insight, we propose a new PEFT scheme called DiffoRA, which enables adaptive adoption of the low-rank decomposition matrices. At the core of DiffoRA lies a Differential Adaptation Matrix (DAM) to determine which module is the most suitable and essential for fine-tuning. We theoretically explain how the designed matrix impacts the convergence rate and generalization capability of a pre-trained model. We then construct the DAM via continuous relaxation and discretization with weight-sharing optimizations. We fully implement DiffoRA and design comprehensive experiments to evaluate its performance. The experimental results demonstrate that DiffoRA delivers state-of-the-art results across multiple benchmarks.

Multimodal Large Language Models (MLLMs) demonstrate strong multimodal understanding capabilities, yet existing prompt-based Multimodal Sentiment Analysis (MSA) approaches largely overlook the modality confidence signals generated during reasoning, leaving models vulnerable when textual and visual cues convey inconsistent sentiment. We propose a Prompt-Level Confidence Modeling (PCM) framework that explicitly extracts modality confidence from structured outputs and integrates confidence-aware fusion with Platt scaling calibration to dynamically assess reliability across modalities. A Prompt-level Modality Disagreement Strategy is further developed to construct conflict subsets, enabling direct evaluation of robustness under cross-modal semantic misalignment. Evaluated on MVSA-S, MSED (English), and CH-SIMSv2.0 (Chinese), few-shot prompting proves most stable overall, with Qwen-VL delivering the best cross-dataset performance—achieving 80.95% Accuracy and 82.22% Macro-F1 on MSED, well above zero-shot and chain-of-thought results. On conflict subsets, calibrated fusion consistently outperforms both direct prediction and uncalibrated confidence aggregation, raising accuracy on MVSA-S conflict samples from 67.61% to 77.53%; ECE and Brier Score metrics further verify that prompt-derived confidence encodes genuine uncertainty information.

Generative models are increasingly used for sequential rec ommendation, but iterative denoising or ODE integration can conflict with low-latency deployment. Recent flow-based recommenders reduce trajectory complexity, and informative history-derived sources have also been explored, yet efficient inference still depends on how the source is constructed and whether the deployed computation is trained di rectly. We present PILOT, a prototype-informed flow-matching frame work for one-step sequential recommendation. PILOT constructs a col laborative prototype prior by fusing a length-robust history summary with a history-conditioned read-out of a shared learnable prototype bank. A FiLM-conditioned Transformer then predicts the target item embed ding from this source. During training, PILOT directly supervises the same t=0 source-to-target branch used at inference, together with a random-time flow-matching branch, next-item ranking, and prior an choring. We use train=deploy only in this computational sense: the de ployed prior, time input, flow network, history condition, and item space prediction all appear in a supervised training branch, while or dinary train/evaluation differences such as dropout remain. On Beauty, Steam, and Yelp, PILOT outperforms all compared sequential and gen erative baselines, including FMRec on every reported metric under three seed evaluation (p < 0.05), while reducing FLOPs by up to about 99% and per-instance latency to roughly 2ms. Code is available at https: //github.com/WentAGan/PILOT.

Financial-service platforms must serve recommendations that are at once engaging and compliant: a high-converting product can also be a money-laundering vector, yet engagement and anti-money-laundering (AML) risk have traditionally been optimized in isolation. We address this gap with a recommender that treats engagement and compliance as a single multi-task objective over heterogeneous user histories. A temporal encoder with time-gated causal attention models the clickstream, a heterogeneous graph encoder reasons over typed transaction graphs, and a gated fusion combines the two before task-specific engagement and risk heads. Training couples a focal engagement loss with a supervised-contrastive risk loss tailored to the extremely sparse AML labels, and a parameter-efficient continual-tuning scheme adapts the model to new products and regulatory regimes without full retraining. On a production financial-service platform with 2.1 billion interactions, the model improves conversion prediction (pCVR log loss 0.0402, an 8.4% relative reduction over the strongest baseline) and AML risk discrimination (AUC-ROC 0.892) jointly under the evaluated operating configuration, and an eight-week online A/B test yields a 12.5% increase in present value under fixed compliance guardrails. Ablations and sensitivity analyses confirm that each component contributes. The results support engagement–compliance co-optimization in regulated recommendation under the evaluated setting.

Sequential recommender systems aim to capture evolving user preferences from behavioral sequences. Frequency-domain methods have recently emerged as an efficient and robust paradigm for sequential recommendation. However, existing models typically rely on data-agnostic bases. This creates a base mismatch: predefined waveforms may not align with heterogeneous behavior patterns, and a single base cannot disentangle mixed intents such as short-term bursts, long-term preferences, and periodic interests. We propose DALW-Rec, a Dual-Adaptive Lattice Wavelet Network for sequential recommendation. To address morphological misalignment, DALW-Rec uses a differentiable lattice structure to learn data-driven wavelet bases while preserving the perfect reconstruction property required by valid wavelet transforms. To overcome the single-view bottleneck, it employs a multi-view frequency filter that learns complementary wavelets in parallel, enabling different views to capture diverse user interests. We further introduce semantic contrastive regularization to align sequences sharing the same target item, improving robustness under sparse and noisy supervision. Experiments on four real-world datasets show that DALW-Rec consistently outperforms state-of-the-art baselines while maintaining linear O(N) complexity.

The integration of Large Language Models (LLMs) as edge- based cognitive services empowers multi-agent systems (e.g., UAV-USV networks) with advanced spatial reasoning. However, offloading inference tasks to untrusted Edge LLMs necessitates transmitting context-rich prompts, exposing sensitive spatiotemporal data to potential inference attacks. While state-of-the-art privacy mechanisms, such as Location Discriminative Geo-Indistinguishability (LDGI), provide snapshot anonymity via continuous numerical noise, they remain vulnerable to belief convergence by physically-aware adversaries during long-duration tracking.To address this, we propose a Privacy-Preserving In-Context Learning (P-ICL) framework. Instead of injecting static noise, P-ICL employs a Semantic Sanitization Mechanism that decouples mission logic from precise physical states via dynamic grid-based tokenization and time-varying permutation. This approach effectively shatters the temporal correlations utilized by adversaries. Extensive simulations demonstrate that P-ICL outperforms LDGI in long-term privacy robustness while maintaining a mission success rate above 94%, as discrete semantic tokens align more naturally with LLM reasoning. Furthermore, achieving a semantic sanitization and reconstruction overhead of only 320 ms, P-ICL provides an order-of-magnitude efficiency gain over cryptographic baselines for real-time edge-cloud coordination.

Effective encrypted traffic classification (ETC) is indispensable for intelligent network management, yet its efficacy is severely undermined by the pervasive long-tailed class imbalance in real-world network environments. Existing methods face two fundamental limitations: generative augmentation often fails to produce reliable and diverse minority-class samples, while conventional classifiers remain dominated by majority classes despite data enhancement. To address these challenges, we propose TFMamba, a Teacher-guided Flow Matching framework leveraging Mamba attention for imbalanced ETC. Specifically, a Flow Matching-based conditional generator synthesizes high-quality and diverse minority-class samples, which are then filtered by a Teacher-guided quality evaluation mechanism that jointly considers distribution consistency, semantic confidence, and decision-boundary margin. Finally, a Mamba-based classifier with modality-specific branches and Focal loss captures long-range dependencies while mitigating residual imbalance. Extensive experiments on three public encrypted traffic datasets demonstrate that TFMamba consistently outperforms state-of-the-art methods, particularly improving minority-class recognition under severe imbalance. It achieves F1 scores of 95.26\%, 99.51\%, and 99.54\% on SJTU-AN21, ISCX-Tor, and ISCX-VPN, respectively. These results confirm TFMamba's ability to generate high-quality and balanced training samples, enabling robust ETC under extreme class imbalance.

The rapid integration of deep learning into web-enabled Intelligent Transportation Systems (ITS) has enhanced automated traffic panel text detection, yet the security and trustworthiness of Transformer-based architectures for Arabic traffic panels remain insufficiently explored. In addition, existing Arabic traffic sign benchmarks present limitations for validating emerging online detection models and continuously updated ITS applications. This paper proposes a security-oriented analytical evaluation of deep neural and Transformer-based models for Arabic traffic panel text detection using the ASAYAR and STP datasets under white-box adversarial attacks (FGSM, PGD, and C&W) and realistic perturbations including fog, Gaussian noise, and motion blur. To improve robustness, we introduce Saliency-Guided Adversarial Feature Regularization (SG-ARFR), a new hybrid defense approach combining adversarial training with Adversarially Robust Feature Regularization (ARFR). Experimental analysis based on Attack Success Rate (ASR), Perturbation Rate (PR), similarity metrics, and feature consistency reveals significant robustness variations across architectures and highlights security risks related to dynamic web-connected ITS environments and online data updates. The findings contribute toward the development of secure, trustworthy, and resilient AI-based vision systems for next-generation ITS applications.

Ethereum fraud detection is challenging because on-chain behaviors are dynamic, heterogeneous, and severely imbalanced. Existing studies have made progress with feature engineering, sequence modeling, and graph neural networks, but many of them focus on a single fraud type, a single modeling view, or static transaction structures, limiting their robustness on realistic Ethereum networks. To address these limitations, we propose TMHG4Eth, namely Temporal Multi-view Heterogeneous Graph Learning for Ethereum fraud detection, a unified temporal multi-view graph framework for detecting Ethereum fraud. We first reconstruct two temporal heterogeneous Ethereum benchmarks by aligning labeled accounts with their surrounding interactions and recovering timestamps and transfer values. We then design a temporal structure-content learning architecture that combines time-aware heterogeneous message passing, structure-content fusion, and auxiliary self-supervision. To alleviate severe class imbalance, we further introduce an interaction-conditioned variational autoencoder (ICVAE) that generates minority fraud nodes and attaches them to the training graph through donor- and similarity-aware edge reuse. Experiments on Ponzi and phishing detection show that TMHG4Eth consistently achieves superior or competitive performance compared with representative feature-based, sequence-based, graph-based, and temporal graph baselines. Additional ablation and generation-quality analyses further verify the effectiveness of temporal modeling, structure-content learning, and ICVAE-based augmentation.

Web-based climate data services increasingly rely on large spatiotemporal data cubes to support interactive extreme-event analysis. A key workload is exact value-to-coordinate retrieval, which identifies the locations of cells that satisfy a value predicate or belong to the top-k values within a query window. Even when analysis-ready Zarr datasets are stored in object storage, these queries remain expensive because every chunk that overlaps the query region must be read and decompressed before its exact contribution can be determined. We present an efficient approach for exact value-to-coordinate retrieval over Zarr-based climate data. Our approach leaves existing analysis-ready Zarr archives unchanged and introduces an external hierarchical metadata index with lightweight per-chunk summaries. These summaries allow non-contributing chunks to be pruned before the underlying data are accessed. For regional top-k queries, the approach further employs an exact, upper-bound-guided best-first search strategy that enables safe early termination. Experiments on ERA5 show that threshold-reporting queries reduce chunk reads by factors of 2.4 to 5.0. Across threshold-reporting and top-k workloads, the approach reduces median end-to-end latency by factors of 2.1 to 6.3, depending on the query-window size, compared with a geometry-only Zarr and Xarray pipeline. Building the initial index requires only 0.47 additional hours of preparation time and 9 GB of extra storage. These results demonstrate that value-aware metadata can provide practical acceleration for exact climate data retrieval while preserving existing file-centric service workflows.

Urban air mobility (UAM) information systems will need to continuously reason over heterogeneous aircraft, vertiports, airspace resources, risk sources, and evolving urban constraints. Existing point-mass or homogeneous traffic abstractions are insufficient because fixed-wing aircraft, eVTOLs, and multirotors occupy airspace with different resource envelopes and induce different risk patterns around corridors, vertiports, population-dense regions, and restricted zones. We propose PI-HSTG, a physics-prior heterogeneous spatiotemporal information graph for maintaining a queryable UAM resource-occupancy state. PI-HSTG encodes aircraft, airspace cells, vertiports, and risk sources as typed nodes, and models spatial adjacency, trajectory occupancy, aircraft competition, risk propagation, and temporal transition as typed relations. Lightweight corridor and funnel priors are integrated with temporal graph learning to support pair-risk, cell-capacity, resource-heatmap, risk-source-impact, and affected-subgraph queries. We develop a reproducible benchmark with hidden wind drift, departure delay, shifted risk sources, unobserved local hazards, OSM-like morphology, and an OSM-derived Hangzhou map layer. Five-seed experiments compare PI-HSTG with homogeneous raw-feature ST-GNN, raw-feature temporal Transformer, and relationattention temporal baselines using conflict PR-AUC, calibration, risk ranking, and occupancy localization metrics. Results show that PI-HSTG is not a universally stronger predictor, but it provides conflict-ranking, occupancy-localization, incremental-update, and cross-morphology riskstate benefits while preserving typed information-system semantics. These findings support physics-prior heterogeneous graph representations as useful predictive state layers for web-enabled UAM information systems.

Audio-Visual Question Answering (AVQA) aims to answer textual questions by jointly understanding visual and auditory information. However, existing methods often overlook subtle motions in complex scenes, limiting their ability to capture fine-grained visual dynamics. Moreover, motion-frequency collaborative modeling remains underexplored, resulting in insufficient cross-modal correspondence. To address these limitations, we propose Q-MFC, a novel method that models fine-grained motion-frequency collaboration from perception to cognition. It begins with a Question-guided Multimodal Interaction (QMI) module, which captures cross-modal interactions and injects question-guided context into visual, audio, and patch-level representations. Next, we design a Motion-Frequency Collaborative Perception (MFCP) module to jointly capture and adaptively fuse semantic visual motions with audio frequency responses. Then, we devise a Multimodal Cognition Reasoning (MCR) module, which integrates local question-guided cues and global audio-visual context to bridge fine-grained perception and cognitive reasoning. Extensive experiments on MUSIC-AVQA and MUSIC-AVQA-R datasets demonstrate that Q-MFC is effective and achieves state-of-the-art performance across a variety of complex reasoning tasks.

Detecting structural anomalies in graphs is crucial to uncover unexpected patterns and mitigate risks in fields such as cybersecurity, social network analysis, and biological research. Existing methods, which often rely on machine learning approaches, struggle to balance detection effectiveness, scalability, and interpretability. In this paper, we introduce RARE (Rarity-based Anomaly Recognition via order Embeddings), an unsupervised framework that combines subgraph mining with Graph Neural Networks (GNNs) to detect rare node neighborhoods indicative of structural anomalies. By embedding subgraphs into an Order Embedding Space (OES), RARE efficiently estimates subgraph frequencies and employs a greedy search to expand infrequent node structures, with interpretable and controllable hyperparameters. Experiments on well- established benchmark datasets demonstrate that RARE consistently outperforms competitive baselines in detection quality and scalability, handling large graphs without memory overhead. It also stays robust where density-based detectors fail and uncovers real fraud structures without supervision. Our implementation and resources are available at https://anonymous.4open.science/r/rare-ad/.

Long-tail metrics can rise by replacing relevant head items with less relevant tail items. We study a frozen recommender whose exposure controller uses only training-observable signals and validation feedback to decide whether to edit a top-K list. SafeTail treats no-op as an explicit action, builds bounded near-tie replacements from host scores and training popularity, and selects a candidate only when validation shows non-inferior accuracy and positive tail movement; its relaxed stage also caps the action rate. Test labels are used once for reporting. Across 3 datasets, 11 hosts, and 99 paired seed runs, SafeTail changes 6.51% of lists, keeps Recall@10 and NDCG@10 near zero change (+0.083% and +0.041%), and improves Coverage@10 by +1.253% and Tail50 by +0.853pp. Negative controls yield larger exposure movement but substantial relevance loss. Thus, long-tail improvement is selective and validation-gated rather than unconstrained coverage maximization.

Cross-domain hashing transfers semantic knowledge from a labeled source domain to an unlabeled target domain, enabling effective large-scale web image retrieval under domain shifts. Existing methods typically rely on cross-domain semantic alignment to learn domain-invariant representations for retrieval. However, the robustness of such semantic alignment against adversarial perturbations remains largely unexplored. Unlike conventional hashing attacks that mainly disrupt Hamming neighborhood structures, we argue that the learned semantic correspondence between source and target domains introduces a critical vulnerability in cross-domain hashing. By slightly perturbing a target domain query toward an incorrect source domain semantic region, the attacker can induce semantic misalignment and produce misleading hash representations, thereby degrading retrieval performance. To investigate this vulnerability, we propose a Semantic Misalignment Attack (SMA) framework for cross-domain hashing retrieval. Rather than merely disrupting hash code relationships, SMA targets the semantic alignment process underlying cross-domain hashing. It constructs stable source domain semantic anchors and uses misleading semantic directions to guide target domain queries toward incorrect cross-domain correspondences. An alignment-induced attack generator is then introduced to produce perturbations conditioned on both visual content and semantic shift information. The generator is optimized with a unified objective that degrades feature-level semantic alignment, disturbs hash-level retrieval relationships, and maintains visual consistency. Experiments on two widely used cross-domain retrieval benchmarks, Office-31 and Office-Home, show that SMA can effectively induce targeted retrieval results under domain shifts, revealing the vulnerability of cross-domain hashing to semantic misalignment attacks. The code will be released upon acceptance to facilitate reproducibility.

Semantic code search and clone detection are essential for software development, maintenance, and reuse. This paper evaluates the effectiveness, efficiency, and scalability of contemporary deep learning models for first-stage recall in large-scale code-to-code search engines. Benchmarking across multiple programming languages and datasets reveals critical limits in the precision and scalability of these models on Terabyte-scale source-code collections. We present LLM-based code normalisation and query-rewriting schemes that yield significant gains in precision for lower-performing models. Our results question the sustainability of resource-constrained deployment and the assumed robustness of current code-specialised LLMs across datasets. We conclude with actionable insights for building scalable, efficient code retrieval systems.

Retrieval-augmented generation (RAG) aims to make language model outputs more accurate and up to date by grounding them in external evidence. However, evidence retrieved from the open web can be outdated, incomplete, or misleading. As a result, a single generated answer may mix well-supported claims with claims that are unsupported or false. This creates a granularity mismatch: rejecting the whole answer discards useful information, whereas keeping it can preserve invalid claims. We introduce Ceres-CRC, a post-generation framework for calibrated claim retention using conformal risk control (CRC). Ceres-CRC decomposes an answer into candidate claims, scores each claim using retrieved evidence, retrieval-free corroboration, and within-answer contradictions, and selects a threshold for a user-specified risk target. Under exchangeability within each calibration cell, and given a fixed scorer and verifier, the calibrated threshold controls the expected fraction of candidate claims that are both retained and labeled invalid. Across three RAG benchmarks with conflicting evidence, Ceres-CRC retains more verifier-valid claims than answer-level rejection and retrieval-only filtering at the same nominal risk target. It also outperforms a retention-matched random baseline and leaves less invalid content than retaining all claims. These results suggest that calibrated post-generation filtering can improve RAG reliability without modifying the retriever or generator.

Long-term time series forecasting requires models that capture global dependencies while keeping computation low. Current frequency-domain methods either operate on the full spectrum or apply a fixed low-pass filter, and therefore do not address how spectral dimensionality can be reduced in an adaptive way. We show that spectra from common benchmarks are highly compressible, but that the degree of compressibility varies across datasets. This observation motivates CRAFT, which learns a complex-valued low-dimensional spectral subspace, makes predictions in that subspace, and uses a residual path to recover information that is not well represented by the compressed branch. Our analysis based on Parseval's identity, best $K$-term approximation, and an optimal linear-subspace characterization supports the compression principle, without assuming that the learned matrix attains the optimum. Experiments on standard forecasting benchmarks and supply-chain demand panels show that CRAFT offers a strong accuracy--efficiency trade-off. Further diagnostics show that its learned projection is more effective than fixed bases, that the frequencies it emphasizes are important for prediction, and that its residual output corresponds to the remaining spectral error.

Modern Web information systems, such as e-commerce, payment, cryptocurrency, and service platforms, generate dynamic multiplex graphs where entities interact through multiple evolving relations. Detecting anomalous nodes is essential for Web security, trust management, fraud detection, and Web mining. Existing methods largely target static attributed graphs, single-relation dynamics, or per-layer reconstruction errors, and thus overlook role-dependent delayed dependencies among heterogeneous relations. Yet a current transaction, access, or communication relation may be explained only by historical behaviors in other relations at specific lags. We propose RCLAD, a role-conditioned relation-lag framework for unsupervised anomalous node detection in dynamic homogeneous multiplex graphs. RCLAD learns relation-specific temporal node representations, infers latent behavioral roles, and estimates sparse role-conditioned kernels that model how lagged source relations support current target relations. It combines structural-attribute reconstruction errors, role-conditioned boundary violations, and lagged prediction residuals into one anomaly score, then applies a role-conditioned conformal FDR rule to obtain labels without anomaly annotations. Experiments on four dynamic multiplex datasets covering cryptocurrency, financial transactions, and network traffic show consistent Macro-F1 gains and competitive or superior AUC over representative baselines.

Multimodal misinformation threatens digital information ecosystems, with cross-modal inconsistency—the divergence between textual claims and visual evidence—as the primary forensic indicator of fraudulent narratives. Existing methods predominantly compress multigranularity representations into domain-agnostic features, or route signals through generic fusion pathways without conditioning on topical context. Two fundamental bottlenecks constrain these paradigms. First, hierarchical conflict context confusion (HCCC) arises from monolithic pooling that erases granularity-specific structural identities, thereby conflating distinct forgery strategies. Second, domain-specific sensitivity heterogeneity persists because identical discrepancy magnitudes carry divergent evidentiary weight across domains—anomalies signaling forgery in politics may constitute post-processing in entertainment—while prior methods apply uniform conflict quantification irrespective of domain context. We propose the Domain-aware Disentangled Interaction Network (DDIN), which reformulates multimodal fake news detection as hierarchical contextual reasoning over multi-granularity conflict evidence. DDIN employs a dual-stream BERT-MAE-CLIP backbone with Multi-Scale Semantic Projection (SSP) decoupling heterogeneous features into shared subspaces, enabling Multi-Granularity Inconsistency Mining (MGIM) operators to extract three matched conflict streams at local-local, global-local, and global-global granularities. A compact Hierarchical Conflict Synergy Network (HCSN) models bidirectional evidence propagation among these streams across granularity layers. Learnable domain gating with alignment-based ambiguity filtering dynamically calibrates conflict sensitivity to each domain's evidentiary standards. Experiments on Weibo, Weibo-21, and FineFake establish SOTA performance, with ablations confirming each component's contribution.

In modern software ecosystems, third-party libraries (TPLs) serve as reusable software services that developers discover and compose. TPL recommendation has attracted growing research attention as a service discovery task in these ecosystems. Graph neural network-based methods model higher-order App-TPL interactions, but fixed propagation gives well-connected libraries more message-passing opportunities than sparsely observed tail libraries, creating a structural visibility gap. To mitigate this gap, this paper proposes TEARec (Tail-aware Edge Augmentation and semantic prototype supervision for third-party library Recommendation). TEARec augments the App-TPL interaction graph with bounded edges for tail TPLs, using App-specific thresholds and per-epoch rebuild from lagged embeddings. A dual-branch architecture encodes nodes through separate similarity-view and interaction-view propagation paths and fuses them via learnable gates, with cross-view contrastive learning aligning head-node representations across the two branches and semantic prototype supervision providing auxiliary regularization from App descriptions. Extensive experiments on a large-scale real-world dataset demonstrate that TEARec improves tail coverage by up to 10.3% (TC@5) over the strongest baseline while maintaining or improving overall ranking accuracy.

Existing multimodal misinformation detection methods often achieve high performance by exploiting easily accessible cues rather than learning genuine evidence of manipulation. Such shortcut learning severely reduces model robustness when data distributions shift across platforms or events. Sentiment is a representative shortcut in the short-video ecosystem, where emotionally charged content frequently exhibits spurious correlations with credibility labels. To address these challenges, we propose CIFD, a counterfactual sentiment debiasing framework for multimodal misinformation detection. CIFD consists of two complementary components. First, a Semantic-Sentiment Hybrid Classification (SHC) module aggregates audience feedback from user comments and integrates it with multimodal content through a sentiment-aware fusion mechanism, enabling the model to capture both content-level and audience-level credibility signals. Second, a Sentiment Counterfactual Debiasing (SCD) module estimates sentiment-induced shortcut effects using a sentiment-only bias estimator and suppresses their influence through principled probabilistic reweighting, thereby forcing the model to rely on genuine multimodal inconsistencies rather than spurious sentiment-label correlations. Experiments on two public benchmark datasets, FakeSV and FakeTT, demonstrate that CIFD consistently outperforms representative multimodal misinformation detection baselines. Additional analyses show that CIFD achieves superior robustness under varying sentiment perturbations and effectively mitigates sentiment-induced prediction biases. These results highlight the importance of explicitly modeling and removing shortcut dependencies for reliable misinformation detection.

Fashion e-commerce platforms represent products using images, descriptions, titles, and structured attributes. Effective cross-modal retrieval over such heterogeneous product information remains challenging because existing methods often treat product text uniformly and rely predominantly on global image-text alignment, leaving visually implicit attributes under-supervised and sentence-level local cues inadequately captured. To address these limitations, we propose FashionSAGE, a fine-grained fashion vision-language pre-training framework. Specifically, (1) FashionSAGE introduces Adaptive Field-Aware Text Masking, with masking probabilities guided by each attribute's visual observability. (2) It incorporates Sentence-Patch Contrastive Alignment, which treats each description sentence as an independent semantic unit and aligns it with selected visual patches via bidirectional asymmetric contrastive learning, thereby capturing sentence-level local cues. On FashionGen, a benchmark for fashion vision-language learning, FashionSAGE achieves Macro-F1 scores of 90.82% and 88.55% for category and subcategory recognition, respectively. In full-gallery retrieval, FashionSAGE achieves 57.96% and 66.71% R@1 on image-to-text and text-to-image retrieval, respectively, while inference-time global-local fusion further improves these results to 58.64% and 68.52%.

In this paper, we propose a parallel computing system for efficient SPARQL query processing over large-scale RDF graphs, leveraging worst-case optimal join algorithms. To reduce redundant computation across tasks, we introduce a locality sensitive hashing (LSH)-based optimization technique. Our approach designs the hash function by incorporating constraints derived from the SPARQL query's property set, to enables intelligent task distribution across multiple computing nodes. This can effectively reduce computational overlap and improve overall execution efficiency. Extensive experiments on both real-world and synthetic datasets show that our method achieves 10\% to 50\% better query performance compared to state-of-the-art approaches.

Graphs model relationships in social networks, web graphs, and knowledge bases. Because storing graphs and processing subgraph matching queries are costly, data owners often outsource graphs to cloud servers. The client must verify the soundness and completeness of returned results. Existing authenticated subgraph matching schemes generate a verification object (VO) along the filtering-searching process. However, they generate verification evidence separately for search branches. When labels have low selectivity or the query contains repeated or symmetric vertices, many partial matchings face the same remaining search, so the same verification evidence is generated repeatedly. Directly merging partial matchings is unsafe under a malicious server: an invalid merge may hide valid embeddings, introduce spurious embeddings, or change the final count. We propose StateCert, an authenticated subgraph matching scheme that makes search-state reuse verifiable. StateCert changes the authentication unit from search branches to canonical states, and represents the folded search as a directed acyclic graph (DAG) of such states, called a Search-DAG. During verification, the client reconstructs each state’s candidate domain and outgoing transitions from authenticated openings, and recomputes the count from leaves to the root. The verified count supports aggregate answers and enumeration cardinality; validity and no-repetition are checked separately. Experiments on real-world graphs show that StateCert reduces median VO size by up to 11.83× and median verification time by up to 7.41× over the strongest baseline. Two verified compressed representations reduce a 12.55 MB low-selectivity answer to 19 KB and 34 KB, and all tested tampering attempts are rejected.

Subgraph matching aims to identify substructures within a data graph that are isomorphic to a query graph. While Graph Neural Network based methods offer promising efficiency, they remain restricted by two challenges. First, scale disparities between graphs cause structural inconsistency: redundant nodes and edges in the data graph pollute the message-passing process, distorting the representations of true matched pairs. Second, existing models lack explicit topological constraints, often yielding structurally invalid predictions. To tackle these issues, we propose an end-to-end framework, Structure-Consistent Subgraph Matching with Gated Cross-Graph Propagation (SC-GCPN). Specifically, our Gated Cross-Graph Propagation (GCGP) module dynamically filters cross-graph information to align node representations and mitigate embedding distortion from complex backgrounds. Furthermore, an Adaptive Structure-Aware Module (ASAM) explicitly enforces neighborhood and feature consistency, imposing rigorous topological constraints to guarantee the structural rationality of matched pairs. Extensive experiments on five benchmark datasets demonstrate that SC-GCPN significantly outperforms state-of-the-art baselines in accuracy while maintaining competitive computational efficiency.

An Integrated Energy Web (IEW) connects oil refueling, gas service, hydrogen supply, electric charging, renewable generation, storage and vehicle-to-grid interaction through regional infrastructure and multi-source energy flows. Most existing studies focus on optimization or scheduling under predefined system structures, while the prior discovery of representative supply patterns from heterogeneous real-world records remains insufficiently explored. This paper proposed a model that represents the Integrated Energy Web (IEW) as an energy-flow heterogeneous graph, in which oil, gas, hydrogen, electricity, and renewable energy records are organized through region-year contexts, energy subtypes, data sources, and facility-related nodes. Supply-pattern discovery is formulated as a coverage-oriented top-k subgraph matching problem. Based on ten IEW supply semantics, we generate query graphs for single-energy supply, joint-energy supply, renewable local production and consumption, and regional multi-energy coexistence. Experiments are conducted on a graph containing 263,532 nodes, 904,691 edges and 258,995 valid energy-flow nodes, and compare Mix-and-Match (M\&M) with LFTJ, VEQ, RM, DSQL and DSQL+. The results showed that M\&M achieves compact and interpretable representative discovery, reaching AC_10=46.82 with a redundancy rate of 19.09\%. These findings indicated that M\&M is suitable for low-redundancy IEW supply-pattern discovery, while VEQ is more effective when larger equivalent result sets are preferred.

Session-based recommender systems (SBRSs) aim to predict the next item based on anonymous, short-term interactions within the current session. However, existing methods often overlook cyclic interaction patterns, which are prevalent in sessions and indicative of deliberation and intent refinement. To address these challenges, we propose Cycle-SR, a novel framework that elevates cycles to first-class modelling primitives and enhances session-based recommendation with cyclic dependencies. Firstly, we extract meaningful simple cycles from intra-session information to construct a cycle-aware graph, which serves as a structural prior. Subsequently, to capture both sequential and cyclic dependencies inherent in the constructed graph, we design a cycle-aware graph neural network (Cycle-GNN) equipped with multi-channel message aggregation, cycle-aware message passing, and cycle-aware session representation learning. Additionally, we leverage contrastive learning to align intra-cycle item representations, and employ non-cycle items within the same session as hard negative samples to enhance discriminability. Extensive experiments on three benchmark datasets demonstrate that Cycle-SR outperforms the state-of-the-art methods.

Community search over heterogeneous information networks (HINs) retrieves a query-centered connected subgraph satisfying semantic type constraints. However, existing methods either assume static snapshots or apply temporal graph neural networks that return evolving embeddings instead of query-contained, type-valid, and connected communities. We formulate continual temporal HIN community search as a query-processing problem whose answer is a community trajectory with persistent, emerging, and disappearing members, and where database constraints query containment, type validity, and connectivity must hold at every snapshot. We propose TDCER, a Temporal Dynamic Community Evolution and Repair framework for continual temporal community search. TDCER: (i) scores candidates with evolution-aware sparse experts and prototype memory for birth, drift, and decay patterns; (ii) maintains the answer incrementally by repairing the previous community only within the changed region R_t of the graph; and (iii) provides a proven validity guarantee on the returned subgraph at every snapshot. TDCER runs in O(r_tkd + |R_t|log|R_t|) per update with r_t≪m_t under local edge updates. On chronological DBLP-Year, ACM-Year, IMDb-TSV, and OGB-MAG workloads under a shared constrained-decoder protocol over ten encoders, TDCER raises future-snapshot F1 by up to 44 points, returns complete community trajectories, and decodes a query in 76.9 ms on an 883.4K-edge OGB-MAG sample, whereas full-score temporal baselines take 0.54--0.56 s.

The widespread adoption of TLS encryption has made traditional detection methods based on ports and plaintext payloads ineffective, leading practitioners to increasingly rely on pre-trained models to identify web applications. Although these models achieve high accuracy, their black-box nature prevents humans from understanding why a data stream is assigned to a particular category, and the low-level attributions produced by XAI methods remain too abstract to interpret, further eroding human trust. To bridge this gap, we propose TrafficXplain, a framework that integrates domain knowledge to guide LLM-generated explanations for encrypted application traffic classification. TrafficXplain combines a Shapley value-based attribution module with a domain knowledge base and human-centered prompting to produce natural and coherent explanations. Experiments on three state-of-the-art pre-trained models show that TrafficXplain generates professional, readable explanations, achieving a composite text-quality score of up to 0.83 and a 7th-grade reading level, while expert evaluation rates its professionalism at 3.78/5, up to 152.5\% higher than ablated variants. Our analysis further reveals that these models may rely too heavily on packet header information rather than the encrypted payload they claim to classify, prompting a reconsideration of trust in both models and datasets. Our knowledge base and supplementary materials are available at \url{https://anonymous.4open.science/r/cdsohw34/}.

Cyber threat intelligence (CTI) is increasingly disseminated through heterogeneous web security reports. Extracting ontology-consistent entities and relations from these sources is essential for threat knowledge graph construction, but strict end-to-end triple extraction remains difficult due to schema ambiguity, long-range dependencies, exposure bias, and large entity-pair candidate spaces. Although large language models show strong extraction ability, they often produce boundary errors and schema-inconsistent triples, while incurring substantial deployment overhead in enterprise security settings. We propose ONSET, an ontology-gated selective state-space framework for joint CTI entity and relation extraction. ONSET combines a shared bidirectional Mamba encoder with lexical and syntactic priors, a sequence-labeling entity decoder, and an ordered-pair relation decoder. An ontology gate enforces type-relation compatibility during relation inference, while an exposure-aware curriculum adapts the decoder to predicted entity noise. On DNRTI-JE, ONSET achieves the best strict end-to-end triple F1 of 74.2%, outperforming the strongest supervised baseline by 3.9 points. On CDTIER, where relation-annotated test labels are unavailable, ONSET achieves the best NER F1 of 91.3%, demonstrating that lightweight ontology-guided extraction remains practical for CTI knowledge graph construction.

Web Application Firewalls (WAFs) are widely deployed to protect web applications from malicious requests. However, SQL injection (SQLi) attacks can still evade WAF inspection through encoding, rewriting, and other obfuscation techniques, making post-WAF traffic auditing essential. While existing SQLi detectors achieve strong performance on raw web traffic, their effectiveness degrades in WAF-allowed traffic. After first-line WAF filtering, the remaining attacks are typically more evasive and expose fewer distinctive attack signals. At the same time, practical post-WAF auditing requires efficient online detection with minimal latency overhead. To address these challenges, we propose ResidualGuard, a lightweight semantic auditing framework for detecting missed SQLi attacks in WAF-allowed traffic. ResidualGuard abstracts payloads into semantic predicates and canonical schemas to capture invariant malicious behaviors beyond lexical obfuscation and learns semantic-aware representations by aligning raw payloads with their semantic abstractions. To improve robustness against difficult benign cases, it further incorporates predicate-guided hard benign samples during training. In deployment, ResidualGuard takes only request text as input to enable fast online auditing with low processing latency. Experiments across multiple WAF systems show that ResidualGuard effectively detects SQLi attacks missed by existing WAFs, demonstrating its practical value as a complementary security layer for deployed WAF systems.

Phishing detectors mix low-cost URL evidence with costly domain, HTML, visual, and reference evidence. Treating these signals as one flat feature vector hides acquisition cost and leakage risk. We present WebPhish-Cascade, a risk-controlled, leakage-aware two-stage cascade with an audited feature taxonomy, URL/domain split protocols, calibrated and learned-benefit routing, and a normalized acquisition-cost model. On UCI PhiUSIIL, containing 235,795 URLs, all-feature models are nearly perfect but use audit-flagged corpus and webpage-size signals. Under leakage-safe domain-group evaluation, a fixed-threshold cascade obtains F1 of 0.9971 at cost 18.77, close to the leakage-safe full model's F1 of 0.9975 at cost 211.00; a stricter risk-budget policy escalates 5.13% and obtains F1 of 0.9970 at cost 27.90. A same-budget learned benefit router further lowers domain-split FNR from 0.0051 to 0.0046. Auxiliary UCI-327, PhishStorm, and SocPhish checks show where the claim does not automatically transfer: UCI-327 requires 85.0% escalation, PhishStorm domain splits reach F1 0.8342, and SocPhish deduplicated-domain URL-only testing drops to AUROC 0.4900. The results support adaptive acquisition as an evaluation and systems pattern, while showing that leakage auditing and dataset-scope limits are essential before interpreting near-perfect phishing accuracy.

TLS has become the default protection mechanism for modern Web communication, but it also enables attackers to conceal malicious activities in encrypted channels. For TLS-encrypted malicious traffic detection, existing spatio-temporal methods struggle to capture malware behavioral characteristics, while TLS-based or behavioral methods often fail to correlate heterogeneous evidence across different feature views. To address these issues, this paper proposes TSTBC, a multi-view detection method for TLS-encrypted malicious Web traffic. TSTBC extracts spatio-temporal statistics, TLS negotiation artifacts, and behavioral connection states from session-level metadata, and integrates them through a gated cross-view fusion mechanism. The gate adaptively emphasizes mutually reinforced evidence while suppressing noisy or weakly relevant features. In addition, we design an evasion-guided feature augmentation strategy to improve robustness against metadata-level manipulation, including statistical mimicry, TLS protocol variation, and behavioral obfuscation. We evaluate TSTBC on nine TLS-encrypted malware families and compare it with seven representative detection methods. Experimental results show that TSTBC achieves strong and balanced performance in terms of accuracy, precision, and F1-score, with an F1-score of 99.1%, outperforming the best baseline by 2.4%.

Encrypted traffic classification is important for network security monitoring. Existing methods have achieved strong performance under fixed class spaces, but practical networks often evolve after deployment, with new applications, protocol variants, and attack behaviors continuously appearing. This requires classifiers to incorporate new traffic classes without repeatedly retraining on historical traffic. This paper proposes ANCHOR, a prototype-guided framework for incremental encrypted traffic classification. ANCHOR learns session representations from multiple traffic views and maintains class knowledge through multi-prototype memory. Newly emerging classes are registered by adding their prototypes, while the representation network remains fixed. To improve deployment robustness, ANCHOR further calibrates prototypes with limited target-domain evidence and performs reliability-weighted prototype matching across feature branches. Experiments on four traffic datasets show that ANCHOR maintains strong and consistent incremental performance compared with representative baselines. Cross-domain and ablation results further demonstrate the effectiveness of prototype calibration and the main design components.

Latent Diffusion Models (LDMs) have brought unprecedented opportunities to AI-Generated Content (AIGC) creation, while also raising concerns over content forgery, copyright infringement, and ambiguous source attribution. Reliable watermarking is therefore important for trustworthy AIGC, as it embeds recoverable signals into generated images for copyright protection and attribution. Existing methods embed watermarks into the latent space of frozen LDM decoders, but frozen-decoder watermarking still suffers from a fundamental compatibility issue: watermark features do not naturally follow the distribution of latent features in the pretrained decoder. We observe that this incompatibility is strongly reflected in spectral orientation mismatch, where inconsistent dominant frequency directions between watermark features and latent features lead to visible watermark-induced artifacts in the generated images. Motivated by this observation, we propose Frequency-Guided Spatial Alignment Watermarking (FSAW), which employs a Spectral Orientation Aligner (SOA) to estimate dominant spectral orientations and align watermark features with latent features before injection. It further introduces a Structure-Aware Fusion (SAF) module to adaptively embed the aligned watermark features according to the spatial structure of current decoder-layer latent features, improving feature-distribution compatibility and stabilizing watermark propagation. Extensive experiments under common image distortions and AIGC-oriented editing attacks demonstrate that FSAW improves both visual fidelity and message recoverability, supporting reliable provenance verification for trustworthy AIGC distribution.

Knowledge Tracing (KT) models students’ evolving knowledge states but is often limited to single-course or pairwise settings, failing to capture multi-course relationships under cold-start scenarios. Moreover, reliance on predefined paths or static graphs restricts generalization. We propose Adaptive Path Evolution for Cross-Course Knowledge Tracing (APEKT), a unified framework for multi-course transfer. A Graph Attention Network (GAT) encodes the cross-course knowledge graph with adaptive edge weighting to model heterogeneous relations. An adaptive path evolution mechanism dynamically explores transfer paths from target concepts by jointly learning expansion and termination, enabling automatic discovery of informative propagation routes. A sequential encoder captures students’ interaction histories to learn time-evolving ability representations, which are integrated with path-based features for prediction. A contrastive learning module further improves robustness by enforcing consistency across node, path, and student representations. Experiments show that APEKT outperforms state-of-the-art methods, especially in cold-start and multi-course settings.

The flexible job-shop scheduling problem (FJSP) is a fundamental NP-hard problem in manufacturing, where scheduling services require both high-quality solutions and low latency. Priority dispatching rules (PDRs) are widely used in practical scheduling systems due to their simplicity and efficiency. However, their predefined priority functions are static and often fail to adapt to dynamic shop-floor states. Deep reinforcement learning (DRL) offers an adaptive alternative, but many DRL-based methods rely on full graph representations and complex graph-to-action architectures, increasing inference overhead and limiting real-time deployment. To bridge efficient rule-based scheduling and adaptive policy learning, we propose a neural priority dispatching rule (NPDR) for FJSP. Instead of mapping graph embeddings directly to action probabilities, NPDR preserves the scoring-and-ranking form of static PDRs by learning a state-dependent neural priority function for feasible operation-machine pairs. For efficient context-aware priority evaluation, NPDR constructs a PDR-oriented decision-frontier representation and incorporates two lightweight context encoders to capture inter-job competition and machine-aware resource assignment information. Experiments on diverse benchmarks demonstrate that NPDR outperforms static PDRs and achieves competitive performance against DRL-based methods while maintaining a favorable quality-efficiency trade-off.

Split Learning (SL) allows resource-constrained devices to collaboratively train the deep neural network (DNN) by only maintaining the shallow layers while offloading the remaining deep layers in the server, which has recently received great attention. However, SL suffers from significant performance degradation when device data features are statistically heterogeneous. Specifically, we theoretically and empirically identify that the root cause of this degradation is the inconsistency in intermediate activation outputs among different devices, making it difficult to share data knowledge among them. Considering this, we propose Feature Alignment based Split Learning (FASL), which aligns the activation distribution across devices such that the heterogeneous data knowledge can be effectively fused. Specifically, FASL first employs a discriminator in the server to distinguish the activation from different devices and in turn provides feedback to train shallow layers in devices to output consistent activation distribution. Furthermore, to enhance the consistency of activation, we introduce the prototypes that represent the activation centroid for each class and train all devices toward prototypes. Analysis shows that our method can theoretically improve the generalization performance as compared to the original splitting methods. Experimental results on various datasets and models show that FASL improves accuracy up to 5.84% compared with state-of-the-art baselines.

Text-to-SQL enables natural language access to relational data in Web information systems. Recent advances increasingly tackle this task through retrieval augmented prompting of Large Language Models (LLMs), where the quality of retrieved demonstrations directly shapes SQL generation. However, existing retrievers often rely on surface level semantic similarity, causing a "semantic-structural mismatch" between retrieved demonstrations and target queries; meanwhile, raw candidate pools introduce substantial structural redundancy. We propose STRIDE, a retrieval framework informed by structure that jointly optimizes pool construction and demonstration retrieval. A two-stage curation strategy removes intra database redundancy and compresses cross database patterns into a compact yet diverse repository, while a module based on distillation predicts coarse SQL sketches to drive a recall-then-rerank pipeline fusing semantic, structural, and schema compatibility signals. Experiments across four Text-to-SQL benchmarks, multiple in-context learning frameworks, and LLM backbones show that STRIDE improves SQL generation performance in most settings, achieves clear gains under lexical perturbation, and reduces the candidate pool size.

An instructor cannot physically monitor the attention of a class of students in an educational environment. Although automated systems have been proposed to address this challenge, they differ significantly in terms of input modalities, performance, and practical applicability. This paper systematically reviews 26 representative studies published between 2016 and 2026. To facilitate a structured analysis, student engagement is examined across three dimensions: behavioral, cognitive, and emotional. We further propose a classification of existing methods based on their input modalities, grouping them into four categories: (1) vision-based action recognition; (2) facial and attention monitoring; (3) multimodal and 3D systems; (4) learning management systems (LMS) log analysis. The analysis reveals that approaches combining multiple data sources and temporal context outperform single-modality methods. However, existing systems are still unable to reliably and simultaneously cover all three engagement dimensions under realistic classroom conditions, highlighting the need for more integrated monitoring solutions.

Automated online textual analysis plays a key role in applications such as social media moderation, online fact verification, and opinion mining, where interpretability has become essential for trustworthy deployment. As an interpretable approach, rationalization provides faithful explanations through a ``select-then-predict'' paradigm that first selects a subset of input tokens as the rationale and then predicts solely from it. Recent studies reveal that rationalization models are vulnerable to adversarial attacks, which can corrupt both predictions and rationales. As a defense, adversarial purification masks and reconstructs tokens to remove adversarial signals without assuming a specific attack form. However, we show that existing purification methods cannot be directly applied to rationalization models: random masking and reconstruction may edit evidence tokens and distort their semantics. Such changes can disrupt rationale selection, leading to degraded task performance and rationale quality even on clean examples. To address this, we propose Selective Purification (SP), which prioritizes non-evidence tokens during purification to reduce unnecessary edits to evidence, and Consistency-Regularized Selective Purification (CSP), which further preserves the downstream model's prediction behavior and token-selection pattern. Experiments on four datasets show that our method better preserves task performance and rationale quality on clean examples while improving robustness under diverse attacks, offering effective protection for interpretable online textual analysis systems.

Grounded long-document fact-checking aims to determine whether a claim is supported by a long source document, requiring fine-grained reasoning over sparse and distributed evidence. Existing methods face a trade-off between verification granularity and computational efficiency. Although graph-based methods improve efficiency through graph summarization, independently summarizing claim and document graphs into global representations may obscure fine-grained evidence and limit cross-graph interaction. Therefore, we propose GLACheck, a multi-granularity cross-graph representation learning framework that captures global semantic consistency, preserves verification-relevant local evidence, and models fine-grained claim--document alignment. The resulting representations are projected into the LLM embedding space as structured soft graph tokens and combined with embeddings of the textual fact-checking prompt for verification by a frozen LLM. Experiments on six public benchmarks show that GLACheck outperforms representative graph-based baselines on average while maintaining computational efficiency. With a frozen 7B backbone, it achieves an average balanced accuracy (BAcc) of 66.9\%, exceeding Qwen2.5-72B (65.9\%) and remaining competitive with other 70B-scale open-source methods.

Generative recommendation retrieves items by autoregressively generating semantic identifiers, a paradigm that implicitly assumes these identifiers already form a suitable language for generation. When semantic IDs are derived solely from text, the resulting code space is often weakly structured and insufficiently discriminative, leaving the generator to decode from an under-structured item language. We argue that images should not be viewed merely as auxiliary features, but as structural priors that help shape the semantic ID space for generation. Based on this view, we propose AVATAR (Aligning Visual And Textual semantics for generAtive Recommendation), a two-stage framework that leverages images as structural priors. AVATAR first performs multi-mask denoising pre-training over multi-modal code blocks to learn a well-structured item language under sparse interaction data. It then aligns textual and visual semantic codes via image-text contrastive learning and conditions generation on the aligned visual semantics through a lightweight prefix image token prompt. Experiments on three real-world datasets demonstrate that AVATAR consistently outperforms strong baselines. Further analysis shows that incorporating images yields more structured and discriminative semantic IDs, corroborating our claim that visual information improves generative recommendation by structuring its code space.

Cross-Modal Hashing (CMH) has garnered widespread attention owing to its efficient storage and retrieval performance. However, existing methods still face two major limitations: 1) most existing methods overlook the importance of label correlation learning for exploiting label semantics; 2) During learning, many methods fail to strengthen label guidance and the learning of relationships among labels. As a result, cross-modal semantic alignment is limited. To address these issues, we propose Label Correlation and Semantic Alignment Learning for Cross-Modal Hashing (LCSA) in this study. LCSA mainly consists of two components, i.e., Global-Local Label Correlation Learning (GLLCL) and Label-Sample Space Alignment (LSSA). GLLCL exploits label correlations at two complementary scales, global and local. The former provides a broad semantic overview, while the latter imposes fine-grained structural constraints. LSSA bridges the label space and the sample space through semantic alignment, so that label correlations can directly inform hash code generation. The formulation is solved through alternating optimization with closed-form subproblems. Experiments on two benchmark datasets demonstrate that LCSA outperforms many state-of-the-art cross-modal hashing methods.

Learned indexes improve lookup efficiency by modeling key distributions, but most existing designs assume stable data distributions and rely on retraining or structural rebuilding when updates accumulate. This limitation is particularly pronounced for string keys, where continuous insertions often cause localized shifts in prefix distributions, degrading prediction accuracy and query stability. To address this problem, we propose ULIS, an updatable learned index for string keys. ULIS introduces a Dynamic Hash Prefix Table (DHPT) to incrementally maintain recent prefix statistics, enabling online correction of short-term and prefix-local distribution shifts without frequent global retraining. It further incorporates a node evolution mechanism that refines local models or expands node capacity according to read/write hotspots. By combining online prefix-distribution correction with selective structural adaptation, ULIS improves robustness under dynamic workloads while preserving long-term prefix modeling stability. Experiments on real-world and synthetic string datasets show that ULIS achieves higher query throughput, lower tail latency, and more stable performance than existing baselines under continuous insertion workloads.

Semantic-ID generative recommenders represent items as discrete code sequences and generate them autoregressively from user histories. Unlearning is challenging because interaction-level requests propagate through sliding-window training data, affecting target-hit and context-hit associations, while dense forgetting updates can perturb the identifier distribution. We propose SCUP-SID, a post hoc approximate unlearning framework that traces request-affected windows, weakens affected associations via reference-anchored preference suppression, concentrates the forgetting signal on gradient-salient coordinates, and preserves retained behavior with a retain-set likelihood term. At inference, prefix-trie decoding enforces catalog-valid generation. Against approximate unlearning baselines on three Amazon datasets, SCUP-SID achieves the strongest deleted-target suppression on Sports and Toys and the membership-inference AUC closest to the retrained reference on all three. On Beauty, it better preserves utility than gradient-ascent baselines, with 1.42% invalid identifiers versus 6.24% for gradient ascent. Ablations show complementary roles of preference suppression and gradient-support localization in balancing forgetting, retained utility, and membership auditing.

Deploying pretrained language models in web-facing applications requires balancing predictive quality against strict memory and latency budgets. Although quantization, structured pruning, and knowledge distillation are widely used to reduce deployment cost, they are rarely evaluated under a common protocol, making their relative trade-offs difficult to assess in practice. We present a unified benchmark that evaluates all three compression families across four web-oriented NLP tasks—news classification, multi-label toxicity detection, extractive question answering, and noisy named-entity recognition—using three encoder architectures (BERT, RoBERTa, DistilBERT). In total, 144 compressed configurations are evaluated against 12 uncompressed baselines under identical hardware conditions. We introduce the Web Compression Efficiency Score (WCES), a single metric that jointly captures quality retention and inference speedup. Our results show that knowledge distillation dominates the accelerated Pareto frontier, yielding up to 7.48× inference speedup with less than 5% quality loss on classification tasks. Quantization is the preferred strategy when memory reduction is the primary constraint but latency slowdowns are acceptable, while post-training structured pruning is consistently dominated by the other two families in this setting. We further demonstrate that compression tolerance varies substantially across task types, with token-level prediction tasks requiring considerably more conservative reduction strategies than document-level classification. Keywords: Model compression · Quantization · Structured pruning · Knowledge distillation · Web-oriented NLP · Inference efficiency

Recent advances in natural language processing have substantially improved task performance, but increasing model complexity has also made predictions harder to interpret. Rationalization improves interpretability by selecting a compact subset of input tokens as a rationale and making predictions solely from it. However, existing rationalization methods are primarily developed for centralized training, making them poorly suited to privacy-sensitive distributed scenarios where data cannot be shared across institutions. Federated learning offers a practical solution that enables collaborative training without exchanging raw data, thereby motivating federated rationalization. A key challenge in this setting is data heterogeneity, which may drive local rationalization models to rely on client-specific shortcuts, thereby compromising both task performance and rationale quality upon aggregation. To address this challenge, we propose Federated Conditional-Independence Rationalization (FedCIR), a federated rationalization framework that learns plausible rationales by enforcing conditional independence between the label and the unselected content given the rationale. FedCIR uses the global predictor's predictions to guide local rationale learning and performs lightweight server-side coordination across clients. Experiments on multiple benchmarks show that FedCIR consistently outperforms baselines in both task performance and rationale quality, converges faster, and remains more robust as data heterogeneity increases.

Fine-grained classification of small web images is challenging because discriminative evidence is often distributed across weak object parts, local textures, and category-specific edge patterns rather than concentrated in one salient region. Recent visual backbones improve recognition through global attention, convolutional modulation, mobile hybrid blocks, or state-space mixing, but these designs are not automatically effective when compressed into a shallow residual classifier for $32\times32$ images. This paper proposes \method, a compact residual classifier built around \block, a residual-subspace feature-fusion unit. Instead of inserting independent modules sequentially, \block first expands the residual channel space into multiple subspaces, then performs progressive residual aggregation among these subspaces, and finally applies lightweight channel calibration to the fused representation. This ordering is essential: channel selection is performed after cross-branch aggregation, so the gate operates on already-composed local, mid-level, and contextual cues. The training protocol, including RandAugment, Cutout, Mixup, label smoothing, staged decay, gradient clipping, and EMA evaluation, is treated as an optimisation layer rather than as a model contribution. On CIFAR-100, \block improves a matched ResNet-18 baseline from $75.87\%$ to $80.15\%$ top-1 accuracy; the complete \method reaches $84.31\%$ top-1 and $96.77\%$ top-5. Protocol-normalised comparisons with eight visual backbones and controlled ablations show that the gain comes from residual-subspace fusion rather than from merely widening the network or appending generic attention. Training dynamics, Grad-CAM, saliency, class-wise accuracy, and confusion analysis further support the same explanation.

User tracking aims to link the same user across multiple web pages and serves as a fundamental capability for enabling personalized services and improving service quality. As dynamic address allocation and encrypted communication reduce the availability of persistent identifiers, traffic analysis-based tracking methods increasingly rely on observable behavior patterns to link user addresses. However, existing methods typically model user behavior as static historical records, overlooking the continuous evolution of Internet environments and limiting their robustness in dynamic scenarios. In this paper, we propose START, a semantic-enhanced temporal graph learning framework for address linkage over encrypted traffic. Specifically, START constructs a semantic-temporal behavior graph from traffic logs, where LLM-generated semantic profiles capture high-level address behavior and temporal edges preserve evolving address-service interactions. START further introduces temporal-invariance and inductive learning objectives to learn stable and discriminative address representations. Finally, START combines these representations with structured pairwise evidence from observable traffic metadata to improve linkage inference. Evaluated on two real-world datasets, START outperforms state-of-the-art methods under static, temporal dynamic, and open-set dynamic protocols. It also achieves strong results in few-shot and cross-dataset settings, demonstrating its robustness for address linkage in dynamic scenarios.

As Large Language Models (LLMs) evolve into the core of Web-based autonomous agents and complex Web Information Systems, their ability to faithfully translate natural language into rigorous structured formats has become paramount, as this capability is critical for Web API invocation and data exchange. However, evaluating this structural fidelity in Web-native payloads remains a challenge: traditional text metrics fail to capture topological consistency in semi-structured Web data, while manual evaluation is prohibitively costly. To address this, we propose Structure-BiEval, a novel self-supervised framework for quantitative, annotation-free assessment tailored for Web data engineering. By leveraging deterministic Intermediate Representations, our framework effectively decouples structure from content, utilizing Content Semantic Accuracy and Normalized Tree Edit Distance as precise metrics. We empirically benchmark 15 state-of-the-art LLMs across dual Web structural topologies, namely Hierarchical Data (Web backend payloads) and Tabular Data (Web frontend presentation). The results reveal substantial variability in structural performance, including cases where mid-sized models unexpectedly outperform larger counterparts in Web data formatting. Furthermore, our findings show that deep recursive nesting poses a consistent challenge for Web agents across varying parameter scales.

We study how public deliberation protocols affect coordination and survival outcomes in an LLM-based multi-agent decision system. We introduce a reproducible simulation framework in which twenty LLM agents, assigned to four personality types (Altruist, Individualist, Rational, and Conformist), vote across five sequential rounds between two options: BLUE, which saves the entire group if it wins the majority, and RED, which guarantees personal survival regardless of the outcome. We manipulate a single between-condition variable (the presence or absence of pre-vote public deliberation) and compare outcomes across one hundred independent runs (fifty per condition). Our results show a dramatic divergence between conditions. Without communication, BLUE coordination collapses after the first round (mean BLUE vote rate of approximately NoCommMeanBlue across rounds; final survivors approximately NoCommFinalAlive out of 20). With communication, agents sustain substantially higher cooperation (mean BLUE vote rate of approximately CommMeanBlue, 95% confidence interval [0.51, 0.73]) and frequently achieve a formal BLUE majority (BLUE win rate of approximately CommBlueWinRate), increasing the number of final survivors to approximately CommFinalAlive out of 20 (Cohen's d = 2.62; Cliff's delta ≈ 1.0). Personality types exhibit behaviorally consistent profiles: Altruists vote BLUE at near-ceiling rates in both conditions (0.70 without communication and 0.99 with communication); Individualists almost exclusively vote RED; Rational agents shift substantially toward BLUE under deliberation; and Conformists show the largest communication effect. These findings demonstrate that even minimal public deliberation can substantially improve coordination in LLM-based collective decision systems, particularly for agents whose behavior depends on perceived group expectations.

Zero-shot 3D visual grounding, which enables open-vocabulary object localization without task-specific training, is fundamentally a multimodal spatial knowledge retrieval and reasoning task. Existing methods often leverage Vision-Language Models (VLMs) as black-box reasoners over rendered multi-view images, implicitly entangling spatial reasoning with visual perception. However, this implicit paradigm obscures the reasoning process and suffers from viewpoint dependency. To address this, we propose ESR, a framework that reformulates the task as explicit spatial reasoning over Perception-Enhanced 3D Scene Graphs, which can be viewed as a structured spatial knowledge graph for embodied environments. Specifically, our method first constructs a perception-enhanced 3D scene graph by integrating the extracted multi-view visual attributes with explicit 3D geometric information. To obtain more reliable spatial edges, we design Size-Adaptive Qualitative Spatial Reasoning (SA-QSR), which accounts for object physical sizes and enables more consistent spatial predicates across objects with different scales. Subsequently, the reasoning process is externalized into First-Order Logic executed over this graph. To improve the efficiency, we design a PEP-Greedy Scheduling strategy that balances predicate pruning power and computational cost. Finally, a VLM-based verification and fallback module is developed to enhance robustness against scene-graph noise, incompleteness, and parsing failures. Extensive experiments on ScanRefer and Nr3D benchmarks demonstrate that ESR outperforms previous SOTA zero-shot methods, achieving notable relative improvements of 9.9\% and 2.5\% in accuracy.

Dysarthria is a motor speech disorder that significantly impairs speech intelligibility and quality of life for millions of people worldwide. Existing digital rehabilitation tools remain inadequate, offering no automated severity assessment, no real-time phoneme-level feedback, and no adaptive personalisation, with Arabic-speaking populations particularly underserved. To address these gaps, we present A’ber, a web-deployed AI-powered speech therapy platform integrating automated severity assessment, phoneme-level error detection, and adaptive therapy planning into a single end-to-end system. The severity classifier, evaluated across self-supervised speech representations including HuBERTLarge and WavLM-Large under strict speaker-independent protocols, achieves 93.33% majority-vote speaker accuracy and 0.8530 quadratic weighted kappa on the UA-Speech corpus under leave-one-speaker-out evaluation, with WavLM-Large yielding the best performance. The error detection pipeline adopts a dual-path architecture combining deterministic goodness-of-pronunciation scoring with an ML-based five-dimension phoneme scorer, fused via Dempster-Shafer evidence theory to deliver clinically safe feedback without committing to incorrect labels under conflicting acoustic evidence. The adaptive therapy engine employs a five-layer architecture combining a Kalman filter for skill tracking, Thompson Sampling for exercise selection, and an Elo rating system for difficulty calibration, continuously personalising a 42-exercise program across five dysarthria-affected speech subsystems. A preliminary usability study with ten participants yielded a System Usability Scale score of 76 with a mean task completion time of 150 seconds, confirming that the system meets all specified non-functional requirements. A’ber represents the first web-deployed AI speech therapy platform targeting dysarthria rehabilitation in the MENA region.

Knowledge graph question answering (KGQA) uses structured knowledge graphs to answer fact-intensive natural language questions. Recent training-free KGQA methods use large language models (LLMs) to explore a knowledge graph step by step. Representative approaches such as Think-on-Graph (ToG) adopt a generative exploration paradigm, where the LLM generates lengthy reasoning text at every exploration step to evaluate and select relations. This paradigm faces two key challenges. First, it is inherently inefficient: candidate relations are evaluated through token-by-token text generation, producing many output tokens at each hop. Second, it can be unreliable for weaker models, which may fail to generate exact relation names or follow strict output formats. We observe that exploration mainly serves to filter relevant paths rather than synthesize answers, and therefore need not rely on free-form generation. Based on this observation, we propose Discriminative Graph Traversal (DGT), a two-stage KGQA framework in which the graph is traversed through independent YES/NO relevance judgments over candidate relations, enabling parallel batched evaluation. The answer determination stage then uses generative reasoning to reason over the retrieved paths and synthesize the answer. DGT is designed to reduce inference time while maintaining or improving answer accuracy. Across the three datasets and open-source backbones, DGT improves over ToG in every controlled comparison, with gains of 3.3 to 36.9 Hits@1 points, while achieving 7.9$\times$ to 16.4$\times$ speedups in average inference time.

Next Point-of-Interest (POI) recommendation plays a crucial role in location-based services and smart tourism, aiming to predict users’ next destinations from their historical check-in trajectories. Recently, many advanced sequence modeling methods based on large language models (LLMs) have been published, but two key challenges remain: i) heavily rely on users’ historical check-in trajectories, causing poor generalization when meeting the out-of-history (OOH) POIs; ii) easily face non-existent or incorrect POIs in recommendations since the inherent hallucination in LLMs. To bridge this gap, this paper proposes a novel LLM-based content-aware multi-view framework for POI recommendation, termed MVF-POI. Specifically, we first leverage an LLM to generate a user preference profile from their historical trajectories, capturing implicit semantic information for recommending OOH POIs, to avoid being hallucinated by directly using the LLM. Next, the temporal-aware trajectory representation learning is designed to integrate POI sequences with time intervals to get temporal dependencies and short-term movement patterns during travel. Then, a multi-view fusion module is employed to fuse semantic and temporal information/features to learn POI representations. Additionally, we introduce SMTR, a novel loss function that explicitly widens the gap between the ground-truth POI and hard negative samples. A series of theoretical analyses and experiments on NYC, CA, and TKY datasets demonstrates the effectiveness and competitive performance of MVF-POI. Code link: https://anonymous.4open.science/r/MVF-POI.

Joint Multimodal Entity-Relation Extraction (JMERE) aims to extract entity-relation quintuples from text with the assistance of visual information, unifying Multimodal Named Entity Recognition (MNER) and Multimodal Relation Extraction (MRE) into a structured information extraction task. Existing studies mainly focus on decoder optimization or external knowledge enhancement, while the effective integration of visual semantics into textual representations remains insufficiently explored. To address this issue, we propose Hierarchical Dual-Granularity Interaction (HDGI), which injects visual semantics into textual representations through encoder-internal hierarchical dual-granularity fusion. Specifically, HDGI employs a Dual Attention Bottleneck module for token-level fine-grained interaction to aggregate critical cross-modal cues and suppress irrelevant visual information, and introduces a Global Semantic Injection Adapter to enhance global contextual semantic modeling. In addition, HDGI adopts MoE-based fine-tuning to improve downstream task adaptation while preserving pretrained representation capability. Experimental results on public JMERE, MNER, and MRE datasets show that HDGI achieves state-of-the-art performance and demonstrates strong generalizability.

Knowledge graphs (KGs) are continuously updated to reflect real-world changes, yet their textual entity descriptions are not automatically synchronized with these updates and become outdated. Existing KG text generation methods mainly produce descriptions from static snapshots; applying them to dynamic KGs often requires regenerating the whole description for local graph changes, introducing redundant inference and unnecessary rewriting. General text editing methods can revise existing descriptions, but they lack explicit control over graph update requirements. To address the problem, we formalize a novel task namely Entity Description Synchronization (EDS), which aims to synchronize existing descriptions with evolved KGs. To tackle the task effectively, we propose GRADE, a synchronization model that treats description maintenance as planned local editing. Specifically, the model uses an attributed neural planner to produce edit plans over existing descriptions, a synchronization engine then realizes and verifies local edits under graph constraints. Experiments on two dynamic KG benchmarks show that GRADE outperforms competitive baselines in synchronization quality. Further analysis confirms that synchronized descriptions provide more reliable textual evidence for downstream entity retrieval.

Knowledge Graph Reasoning (KGR) is an efficient approach to relieve the incompleteness and sparsity problems. Due to the inherent heterogeneity of KG, although the Graph Neural Network (GNN)-based KGR models can achieve more optimal performance recently, they have neglected the correlation between homogeneous entities/relations and still suffer from the problem of insufficient feature information acquisition. This paper proposes an Adjustable lower-heterogeneous granularity subgraph Decomposition-promoted Graph ATtention mechanism for KGR, namely ADGAT. It can effectively reduce reasoning noises caused by the heterogeneity of KG, distinguish the contribution degree of entities and relations at different levels, while granting the mechanism's flexibility. Extensive experimental results and ablation studies on four KGs explicitly demonstrate significant improvement of the proposed ADGAT mechanism compared to current GNN-based methods.

Temporal knowledge graph (TKG) extrapolation aims to predict future facts from historical interactions. Existing embedding-based methods rely on entity-specific historical signals and often produce unreliable representations for sparse entities, while semantics-based methods may suffer from incomplete retrieved contexts, frequency bias, and unconstrained entity generation. To address these limitations, we propose PAST, a semantics-based framework that reformulates entity prediction as semantic ID (SID) generation. PAST constructs hierarchical SIDs via residual quantization over description-enriched entity representations, enabling semantically related entities to share prefix tokens. It then adapts an LLM to generate target SIDs through two-stage instruction tuning. During inference, PAST performs prefix-constrained hybrid decoding over an SID trie, where embedding-based entity scores are projected onto valid prefix branches and fused with LLM token probabilities to preserve plausible sparse-entity prefixes. Extensive experiments on three datasets demonstrate that PAST outperforms state-of-the-art methods and is particularly effective in predicting sparse entities.

Knowledge Graphs (KGs) have grown too large to train embedding models on a single machine. Distributed training solves this problem by splitting the KG across multiple workers, but its success depends entirely on how the graph is partitioned. Traditional partitioning methods like METIS and KaHIP treat all relations as identical. They cut semantically important connections, which forces workers to communicate constantly and degrades model accuracy. We propose SAKGP, a semantic-aware knowledge graph partitioning method. SAKGP computes three complementary scores for each relation type: structural centrality (RCS), semantic specificity via entropy (RSS), and local triangle density (RTS). These scores are normalized and combined into an adaptive weight using differential evolution. A Semantic Entity Partition (SEP) score then places entities into partitions by balancing intra-partition affinity against inter-partition leakage, followed by iterative refinement with Semantic-Aware Label Propagation (SALP). We evaluate SAKGP on three benchmarks using the KGTransformer model in a distributed environment with up to 16 workers. Compared to METIS and KaHIP, SAKGP reduces edge cuts by 4-10 percentage points and improves MRR by 6-8% in distributed settings. With 16 workers, SAKGP retains 93% of centralized accuracy while achieving near-linear speedup up to 12 workers.

Quantifying node importance in social networks is essential for identifying effective intervention targets. Traditional centrality measures rely mainly on static topology and do not directly answer which nodes are valuable to boost through external incentives. In this paper, we study source-boosting node importance under the Independent Cascade model. We define the Global Boosting Score (GBS) of a node as its expected marginal increase in influence spread when the initial seed is drawn uniformly at random from the network. To estimate GBS efficiently, we develop the Global Node Importance by Boosting (GNIB) estimator based on three-state reverse sampling. For each sampled root, the estimator counts the single-node seeds that can newly reach the root after boosting a candidate node. We prove that the estimator is unbiased for the defined GBS objective and parallelize it across independent samples. Experiments on seven real-world networks show that GNIB is orders of magnitude faster than a direct Monte Carlo baseline while closely matching the MC reference rankings, scales near-linearly with the number of threads, and yields scores that correlate positively with real user engagement.

Heterogeneous information networks (HINs) encode diverse semantics through complex cross-type relations. Meta-structures are powerful tools for capturing such semantics, yet existing methods often treat them as isolated designs and overlook their inherent partial-order relationships. In this paper, we propose a principled framework grounded in order theory that defines and computes partial-order relations among meta-structures, organizing them into a partial-order meta-structure hierarchy for semantically consistent structure selection and composition. Building upon this hierarchy, we develop the Partial-Order-based Meta-structure Representation Learning Network (POMRLN), which maps meta-structures to structure-induced propagation operators and performs convolution through a positive semidefinite (PSD) polynomial filter to stabilize semantics-aware high-order propagation. Extensive experiments on six real-world HINs demonstrate that POMRLN consistently outperforms state-of-the-art baselines on both semi-supervised node classification and link prediction, confirming that partial-order-guided organization provides an effective framework for extracting meaningful semantic relationships in heterogeneous graphs.

In querying relational layouts of large RDF datasets at scale, relatively little attention has been given to the performance impact of pre-execution runtime prediction and to recommending an RDF relational layout before the workload is materialised and executed. This gap is important because deciding the best layout is not straightforward: it depends on the structure of the queries, the mixture of query shapes in the workload, the predicate distribution in the dataset, and the execution behaviour of the underlying SQL Big Data engine. This paper addresses this gap by separating performance modelling from layout recommendation, considering both single-query and workload-level performance prediction, and formulating RDF relational layout recommendation as a workload-level score-based ranking problem. Unlike the purely descriptive benchmarking studies, which constitute the majority of the literature, our proposed framework uses static SPARQL algebra features, dataset/layout statistics, and workload profiles to support query-level runtime prediction, workload-level runtime prediction, and score-based layout recommendation before full execution. Our experimental results show that Ridge regression, the best-performing recommendation method, achieves pairwise layout-ranking accuracy of 0.79 and mean regret of 0.041, indicating that even when the workload-aware layout itself is not always the fastest layout, score-based ranking can select near-optimal layouts.

Website Fingerprinting (WF) commonly assumes that website visits start with an empty browser cache. In practice, cached Web resources expire, validate, and reload at different rates, causing encrypted traffic to drift continuously with the elapsed time between visits. We study this temporal cache shift through a Web measurement of 7.14 million resources across 80,000 websites and show that discrete cached snapshots cannot adequately cover realistic revisits. We introduce Cache-Time Snapshot Cloning (CTSC), which efficiently reproduces browser cache states at specified elapsed intervals without waiting for the corresponding wall-clock time. We further propose \textsc{CacheTraj}, which learns temporally coherent traffic embeddings and represents each website as a piecewise-linear cache-time trajectory. An offline LLM-encoded cache-semantic gallery aligns resource-level cache behavior with traffic prototypes and reranks trajectory candidates, while deployment requires only encrypted traffic. Experiments on 640 websites demonstrate strong robustness at held-out cache intervals in closed-world and open-world settings, as well as across website collections and geographic locations. These results show that modeling the temporal semantics of Web caching substantially improves WF under realistic browsing conditions.

Secure voice-based authentication systems rely on audio deepfake detection modules to verify the authenticity of user speech before or alongside speaker verification. However, these modules may suffer from shortcut learning during deployment shift, where models exploit non-causal artifacts rather than spoofing-related evidence. Existing studies mainly attribute this issue to dataset-dependent artifacts and training-distribution bias, while feature-level and architecture-related shortcut risks remain less explicitly analyzed. This work analyzes shortcut learning from these two perspectives and investigates the corresponding mitigation strategies: low-frequency Sinc-band feature constraint and information-preserving local-global architecture design. These strategies are instantiated with a Sinc-band-4 kHz front-end and a Rawformer-style architecture, respectively. Experiments on ASVspoof 2019 LA and ASVspoof 2021 LA show that the proposed mitigation strategies can improve out-of-distribution generalization under standard evaluation. We further conduct asymmetric spurious-cue stress tests on ASVspoof 2019 LA and ASVspoof 2021 LA/DF with MP3 compression and additive white noise, where cue-label associations are reversed between training and testing. Results show that the Sinc-band-4 kHz front-end provides stable gains, while Rawformer and Light-Rawformer offer additional but less consistent improvements. These findings also suggest the complementary roles of the feature constraint and architecture design in reducing shortcut reliance for trustworthy voice authentication.

The growing volume of spatial datasets and the need for efficient search pose significant challenges in both privacy preservation and efficiency. To address these challenges, we propose efficient and privacy-preserving exemplar-based spatial dataset search schemes, namely EPES and EPES+. We first introduce a grid density distribution-based similarity model (GD-distribution), which converts spatial datasets into density vectors to support secure similarity computation without exposing raw data. Based on this model, we develop a baseline scheme, EPES. To further enhance search efficiency, we construct a spatial dataset grid inverted index (SGI-index) and propose an optimized scheme, EPES+, which significantly prunes the candidate set before similarity computation. Comprehensive security analysis and experiments demonstrate that both schemes achieve high accuracy, while EPES+ achieves up to 6$\times$ speedup over PriDAS+, enabling secure and practical exemplar-based spatial dataset search in cloud environments.

Machine learning (ML)-based network intrusion detection systems (NIDSs) are increasingly used to protect Web-facing infrastructures, but they remain vulnerable to realizable packet-level adversarial attacks. Existing adversarial defenses mainly focus on feature-level adversarial attacks, where defenses learn from or detect perturbation-induced patterns in adversarial feature samples. However, packet-level adversarial attacks modify real packet sequences through functionality-preserving operations rather than directly perturbing features. These constraints make their feature patterns differ from perturbation-based adversarial feature samples, limiting the effectiveness of existing defenses. To address this problem, we propose ECLIPSE, a defense framework for detecting packet-level adversarial traffic against ML-based NIDS. ECLIPSE constructs operation-stable local views from raw flows, learns persistent malicious-core evidence through complementary deep and structural survival branches, and detects adversarial traffic by measuring the inconsistency between the base NIDS output and the learned survival evidence. Experiments on three real-world datasets, multiple victim detectors, and four packet-level attacks show that ECLIPSE achieves an average Adversarial Detection Rate (ADR) of 0.922 and an F1-score of 0.911, outperforming state-of-the-art defenses. Robustness and ablation studies further confirm the effectiveness of its key components.

Multimodal sequential recommendation incorporates textual and visual information to alleviate the limited semantic expressiveness of traditional ID-based methods. However, most existing approaches fuse ID and multimodal representations at an early stage, ignoring their heterogeneous roles and causing cross-signal interference. Moreover, they often insufficiently model how the importance of different signals evolves with irregular interaction intervals. To address these issues, we propose TDMIRec, a progressive decoupling framework that follows a decoupling--coordination--enhancement paradigm. Specifically, TDMIRec separates user preferences into two role-specific channels: an ID-based commonality channel for collaborative behavioral patterns and a multimodal individuality channel for content-level semantic preferences. We introduce a time-aware dynamic coordination mechanism to adaptively adjust the contribution of each channel at different time steps using observed time intervals. Finally, a set-aligned multi-interest module is designed to capture multiple coexisting user interests and enhance fine-grained preference matching. Experiments on four real-world datasets show that TDMIRec consistently outperforms state-of-the-art baselines, demonstrating the effectiveness of the proposed framework.

Multimodal technologies and causal inference are widely used in recommendation tasks and have yielded substantial progress. However, existing studies still suffer from two key limitations. From a temporal perspective, most multimodal causal models treat user interests and conformity effects as static factors, ignoring their dynamic evolution along interaction sequences and their time-varying causal influence on sequential behavior. From a feature perspective, causal sequential methods are mainly built on ID-based representations, lacking semantic awareness and the ability to capture popularity-inducing cues in multimodal content. Although conventional multimodal methods introduce visual and textual features, their coarse-grained fusion often entangles intrinsic item semantics with external popularity signals, causing models to mistake conformity-driven interactions for genuine preferences. To address these issues, we propose Dynamic Causal Intervention Multimodal Sequential Recommendation (DCIMSR). We first design a multimodal feature disentanglement module to separate intrinsic semantic features from popularity-induced features in visual and textual modalities, providing fine-grained observed variables for causal modeling. We then construct a dynamic multimodal causal graph that incorporates temporal evolution, popularity-induced cues, and instantaneous item popularity to estimate time-varying conformity effects. Based on this formulation, DCIMSR performs sequence-level causal intervention with a dual-tower sequential architecture, suppressing conformity interference while preserving users’ genuine temporal preferences. Extensive experiments on multiple public datasets show that DCIMSR consistently outperforms representative baselines, demonstrating the effectiveness of fine-grained multimodal disentanglement and dynamic conformity modeling. The source code is available at https://github.com/PiBigStar589/DCIMSR.

Academic paper recommendation plays an important role in helping researchers discover relevant literature in scholarly information systems. However, user--paper interactions are often highly sparse, making it difficult for graph-based recommender models to learn reliable representations for newly published, long-tail, and low-interaction papers. Although auxiliary academic relations such as paper similarity and author collaboration can provide complementary signals, these proxy structures are not always reliable and may introduce noisy or weakly related neighbors into the representation learning process. To address this problem, we propose SCUFRec, an uncertainty-aware proxy structure completion framework for sparse academic paper recommendation. SCUFRec first constructs dual proxy structures for low-interaction papers by aggregating pre-computed semantic-similarity neighbors and co-author neighbors. It then introduces a confidence-driven fusion and gated update mechanism to adaptively combine the original collaborative representation with the completed proxy representation. To further reduce the risk of noise propagation, SCUFRec employs multi-expert probabilistic modeling to estimate sample-level uncertainty and uses a trust gate to control the contribution of proxy-enhanced representations. In addition, local consistency alignment and cross-view contrastive regularization are incorporated to improve representation robustness. Experiments on four academic recommendation datasets, including AMiner-WeChat, AMiner-PC, CiteULike, and DBLPV12.ai, show that SCUFRec consistently improves Top-K recommendation performance over representative baselines, with more evident gains for cold-start and low-interaction papers.

The dark web has become a major hub for exchanging information, services, and resources related to criminal hacking. Discovering these activities is critical for cyber-threat intelligence (CTI), yet remains inefficient, ineffective, and potentially harmful to researchers. To address these limitations, we investigate the automatic discovery of hacking resources from recommendations shared within dark-web communities. Our approach leverages explicit Tor and I2P links and site-name mentions embedded in forum discussions as starting points for resource discovery, then enriches each recommendation with evidence from the seed post, neighborhood discussions, landing page, search-engine results, and site descriptions for resource classification. Using a curated multilingual dataset, we compare preprocessing strategies, representations, and learning paradigms ranging from lexical and transformer embeddings to fine-tuned LLMs, obtaining the highest F1-score of 0.839 when all five evidence sources are integrated. Unexpectedly, this configuration combines basic text preprocessing, dense lexical representations, and a lightweight neural architecture to outperform more complex semantic and LLM-based alternatives. These findings suggest that combining complementary recommendation signals matters more than adopting complex representations or learning models for discovering CTI resources.

Resource-constrained devices increasingly require hybrid vector retrieval that combines approximate nearest neighbor search with structured filtering. However, existing retrieval systems are primarily designed for memory-rich and relatively static environments, resulting in high memory overhead, insertion cost, and query latency when supporting dynamic workloads under limited resources. This paper presents HIVF, a low-overhead hierarchical inverted file indexing framework for resource-constrained hybrid vector retrieval. HIVF organizes the vector space using a two-level clustering structure, which reduces centroid matching cost and the associated indexing overhead. To limit memory consumption and disk access, HIVF separates lightweight routing metadata from disk-resident posting lists and employs memory-mapped storage with LRU-based cache management. To support online insertions, HIVF combines an incremental update buffer with a query-throughput-aware rebuild strategy, which determines when buffered updates should be merged into the main index according to workload conditions. For post-filter hybrid queries, HIVF further employs a selectivity-aware probing strategy that adapts the search scope to filtering selectivity and applies recall compensation when the initial search produces insufficient valid candidates. Extensive experiments on SIFT, GIST, and DEEP10M demonstrate that HIVF reduces index construction overhead, memory usage, and insertion cost compared with a flat IVF baseline, while achieving competitive retrieval accuracy. The adaptive probing strategy further provides a favorable latency-recall trade-off across different filtering selectivities. Overall, HIVF offers an effective approach to hybrid vector retrieval under resource-constrained and dynamically evolving workloads.

Retrieval-augmented generation has evolved from text-only corpora to heterogeneous evidence environments encompassing passages, tables, and knowledge graphs. However, once retrieval recalls cross-modal candidates, how to assemble them into a context that preserves complementary evidence without redundancy remains a neglected bottleneck. We frame this assembly as a combinatorial optimization problem, Multimodal Evidence Composition, and prove its NP-hardness via reduction from budgeted maximum coverage. We derive sufficient conditions under which greedy selection achieves a constant-factor approximation guarantee. Building on this, we introduce UniFusion, an entropy-adaptive controller that unifies query routing, cross-modal reranking, and budgeted packing. UniFusion leverages a single uncertainty signal to dynamically balance dense-sparse fusion and enforce soft modality quotas, jointly preventing modality collapse and suppressing redundancy. Experiments on the mmRAG benchmark show that UniFusion outperforms various baselines in both retrieval and generation. Ablations reveal a retrieval-generation paradox: removing redundancy penalties improves ranking metrics while degrading generation performance, confirming that better retrieval does not automatically translate into better reader contexts. Effective heterogeneous context assembly thus requires jointly optimizing modality balancing and redundancy suppression.

Dense retrieval is a core mechanism for Web information systems, enterprise search, and retrieval-augmented generation. In heterogeneous Web corpora, however, most dense-retrieval pipelines search a single global vector space in which all information units compete for every query. This flat organization can keep topically irrelevant regions active, increasing access cost and admitting contextually mismatched evidence. We propose topic-guided semantic space partitioning as a lightweight retrieval-control layer for elastic access to heterogeneous Web information. The method induces overlapping topical subspaces offline and uses query-topic relevance online to activate a query-specific candidate space before dense scoring and policy-specific ranking. Because it preserves the encoder and similarity function, the contribution is candidate-space control rather than a new relevance model. We evaluate the framework in retrieval-augmented generation settings with public question-answering benchmarks, an additional labelled setting, and anonymized corpus diagnostics. Across complete latency comparisons, hard topic filtering reduces per-query retrieval latency by 4.4 - 19.8% relative to global dense retrieval. The reported answer-quality metrics remain comparable to the global baseline, with positive BLEU-1 and recall-weighted ROUGE-L deltas in the available quality settings. Topic-induction diagnostics further show that partition granularity affects coverage and should be selected empirically. Overall, the results support explicit topical organization as a practical, interpretable control layer for scalable Web information access.

Cross-modal hashing (CMH) has emerged as an effective technique for retrieving relevant data across different modalities, attracting increasing attention due to its low storage costs and fast query speeds. Although existing CMH methods have achieved remarkable progress, most approach the problem by preserving pairwise similarities, often neglecting the intrinsic geometric structure and the latent linear correlations within categories. Based on this observation, we posit that samples belonging to the same category should lie in a low-dimensional subspace (low-rank), while the holistic data distribution within a modality should remain diverse (high-rank). To address this, we propose a novel Integrated Low-Rank Hashing (ILRH) method for cross-modal retrieval. Uniquely, our approach integrates label information as an explicit third modality within a unified optimization framework. Specifically, ILRH enforces the low-rank property on intra-class samples while simultaneously maintaining the high-rank property of the global view. This strategy effectively aligns heterogeneous modalities while enforcing a strict subspace structure on the learned hash codes. Extensive experiments on three widely used benchmark datasets demonstrate the effectiveness and robustness of the proposed ILRH over state-of-the-art CMH approaches.

Conversational agents typically present the same response style to every user. When a chatbot cannot do what a user asks, frustration is often unavoidable — but the way the limitation is communicated is a design choice, and we ask whether adjusting it reduces frustration. In a vignette-rating experiment (N~=~235), participants rated frustration and trust over chatbot responses to a fixed, frustrating scenario that varied along length, uncertainty, and formality. We estimated population-level effects with linear mixed-effects models and tested whether individual differences in form sensitivity form discrete user profiles or vary continuously. We found that longer and more formal responses reduced frustration while uncertainty had only a small effect; that trust moved through the same two dimensions in the corresponding direction and was unaffected by uncertainty; and that frustration and trust were strongly and negatively coupled within participants (r = -.80). Sensitivity to the three dimensions studied varied continuously and independently rather than forming discrete types, and was barely predictable from individual differences: most participants were less frustrated by more formal responses, length divided them (a majority less frustrated by longer responses, a substantial minority more), and uncertainty barely registered. These results suggest that adjusting response form --- length above all --- is a low-cost lever on frustration, but that an adaptive agent must tune it to each user individually rather than to a fixed type.

Operational logs are increasingly consumed by AI-assisted troubleshooting agents, creating an indirect prompt-injection risk: adversarial text hidden in otherwise normal log fields can later steer a model that reads those logs. We treat its detection as unsupervised anomaly detection and build a detector from a single Growing Self-Organizing Map (GSOM) trained only on baseline logs. Each log line is reduced to a format-agnostic payload and encoded as a feature vector that combines a pretrained sentence embedding with lexical, structural, and reference-anchor signals; the GSOM summarizes normal behavior as a few hundred prototypes, and an event is flagged when its distance to the nearest prototype exceeds a threshold calibrated on baseline data. We evaluate on multiple public log sources, each carrying injections in its own attacker-controlled field across a range of prompt-injection categories with adversarial hard negatives. Against a set of unsupervised baselines, the GSOM achieves consistently high detection accuracy on every source and stay accurate across all of them while being far more compact and well suited to streaming.

Cost-based query optimizers depend on accurate cardinality estimates, but for spatial predicates every estimator used in practice—spatial histograms, R-tree summaries, and the statistics shipped with production engines—reasons only over minimum bounding rectangles (MBRs). Such estimators predict the filter cardinality and ignore the refinement step that verifies exact geometry; for the irregular polygons common in real data the two differ by up to an order of magnitude, so optimizers systematically mis-cost their plans. We argue that closing this gap requires modeling the shape of the query region, and we show that a classical signal-processing tool—the Fourier descriptors of the query contour—does so within the optimizer’s time budget: the boundary is encoded as a periodic complex signal whose leading spectral coefficients form a fixed-length feature that augments the MBR and feeds any off-the-shelf regressor. On the YAGO2 knowledge base (≈4.5 M point objects), Fourier-based estimators reduce the true-shape error by nearly 5× over the strongest non-sampling baseline (9,936 vs. 48,305 mean absolute error) and match the accuracy of geometry-aware sampling while running five orders of magnitude faster—sub-millisecond per estimate versus ∼80 ms—at a cost of only a few megabytes of memory. We further characterize when an estimator is fast enough to be invoked during plan enumeration, and show that shape-aware learning is the only approach that is at once accurate on refinement and practical at optimization time.

Temporal Knowledge Graph (TKG) extrapolation aims to predict future facts from historical observations and plays an important role in event forecasting and risk analysis. Recent virtual-graph reasoning methods enhance extrapolation by first constructing a virtual subgraph from preliminary predictions and then performing a second round of graph reasoning. However, these methods make limited use of relational semantics. Low-frequency entities often learn unreliable dynamic representations due to sparse historical interactions, while virtual subgraphs are constructed mainly according to prediction scores, which may introduce semantically implausible candidate edges. To address these limitations, we propose SEER, a lightweight semantic-enhancement framework for virtual-graph-based TKG extrapolation. SEER constructs a shared cross-relation semantic operator by jointly modeling entity--relation memberships, entity frequencies, and relation-transition patterns, and applies it at two complementary stages. Specifically, Cross-Relation Semantic Compensation (CRSC) enhances second-pass entity representations with relation-role semantics, while Semantic Edge Screening (SES) re-ranks candidate edges before virtual-subgraph construction using semantic compatibility. The two components share the same semantic operator while introducing only limited additional computation. Experiments on three ICEWS benchmarks demonstrate that SEER consistently improves the performance of a strong virtual-graph reasoning backbone. Further ablation studies, frequency-stratified analysis, and virtual-graph diagnostics show that SEER is particularly effective for low-frequency entities and constructs higher-quality virtual subgraphs.