Accurately extracting entity relationships from text is challenging due to information redundancy, and errors can propagate misinformation with security implications. Document-level relation extraction (RE) presents a particular challenge due to its inherent complexity, requiring the synthesis of entity relationships distributed across multiple sentences within a document. Existing methods often rely on simply stacking network layers to address long-range dependencies, which can result in insufficient model robustness and errors when processing lengthy contexts. To address this, we propose RLNA-Net, a novel method that reframes document-level RE as a semantic segmentation task. Our model integrates a ResNet architecture with Layer Normalization to stabilize training and mitigate degradation caused by deep networks. Furthermore, we incorporate Multi-Head Self-Attention (MHSA) into the backbone to effectively capture diverse and long-range dependencies within the document. Comprehensive experiments on three benchmark datasets—DocRED, CDR, and GDA—demonstrate that RLNA-Net outperforms existing strong baseline models.
Federated learning is a distributed learning paradigm, where each client jointly trains a global model by only sharing gradients. Nevertheless, previous research has shown that the shared gradients can be used to privacy leakage, i.e., the so-called gradient inversion attack. Notably, existing attack studies have some generic shortcomings, such as insufficient theoretical guarantee and poor reconstruction when given difficult conditions. In our work, based on variational Bayesian theory, we propose a gradient inversion attack combined with the novel diffusion models, exploiting its remarkable expressiveness of image manifold to improve the inverted results. Compared with existing research, we provide elaborate theoretical analysis for our attack, showing its good adaptability in face of two real-world attack scenarios, namely as sample-level and user-level inversion. Through comprehensive validations under diverse datasets and realistic circumstances, we hope that our work can help understand and evaluate such privacy risk in federated learning.
As large language models (LLMs) scale, their inference-time computational cost and latency become increasingly prohibitive. Early-exit (EE) methodologies mitigate this issue by reducing the depth of effective inference, terminating the forward pass once the confidence in the prediction of the model exceeds a predefined threshold. However, existing EE approaches typically employ static thresholds, which can result in either premature exits that degrade predictive performance or unnecessarily deep computation that diminishes efficiency gains. To overcome these limitations, we propose DATEE, a Dynamic Adaptively Thresholded Early Exit framework based on a hierarchical trend sampling strategy. First, we develop a hierarchical trend-sampling mechanism. The first-stage sampling is conducted at shallow network layers to assess whether a predefined confidence threshold is met, while incurring minimal computational overhead. The second-stage sampling is then executed at deeper layers to capture and characterize confidence stabilization patterns. Second, we introduce a dynamically adaptive threshold method based on sampled feature representations, which jointly optimizes the confidence threshold and the initial sampling layer, thereby enabling coordinated and autonomous self-optimization. Experimental results on diverse natural language processing benchmarks show that DATEE stabilizes exits and improves inference efficiency by 1.1 × –4.9 × . Furthermore, it substantially reduces the average depth of inference and computational cost while preserving overall model performance.
The extensive generative capabilities of Large Language Models (LLMs) are causing the potential to increase abuse and misuse of AI-generated contents, which makes annotation and detection be critical tasks. Embedding watermarks into models is considered an effective approach that is imperceptible to humans but can be easily detected using specific algorithms. However, compared to the image and audio domains, the text domain possesses relatively limited redundancy. Most existing watermarking methods struggle to be effectively applied in scenarios requiring substantial embedding space, including low-entropy generation tasks and when it is necessary to embed multi-bit watermarks for advanced traceability. To address this challenge, in this work we propose a novel entropy-aware multi-bit watermarking framework. Specifically, by developing an entropy-aware filtering module, we estimate context entropy in real time to filter out invalid watermark spaces. Simultaneously, we integrate a multi-bit position allocation strategy to efficiently utilize the remaining space, thereby enabling trustworthy traceability of multi-bit watermark information in diverse generation scenarios. Comprehensive experiments demonstrate that our scheme improves text quality across various generation scenarios while maintaining high multi-bit decoding accuracy, providing a practical solution for the trustworthy traceability of LLMs.
Multi-agent reinforcement learning (MARL) faces significant challenges in unshaped scenarios where the state and action spaces dynamically change due to agent deaths or other environmental factors. Traditional methods either ignore these changes or employ hard masking techniques, leading to unstable learning and suboptimal policies. To address these limitations, we propose DP-QMIX, a novel approach that integrates a Death Prediction Model with the QMIX framework. Our method introduces three key innovations: (1) a multi-task DP-Model that jointly predicts agent survival probabilities and next states using LSTM networks, enabling progressive decay of dying agents’ action-values; (2) a self-weighted mixing network architecture that dynamically adjusts agent contributions based on survival probabilities; and (3) a soft elimination mechanism that stabilizes the learning environment by smoothly transitioning state spaces. Extensive experiments on StarCraft II micromanagement scenarios demonstrate that DP-QMIX achieves faster convergence (approximately 28 https://github.com/Vergissmeinncht/DP-QMIX .
Cross-domain recommendation (CDR) improves target-domain recommendation by leveraging behavioral signals from auxiliary domains, alleviating data sparsity and cold-start issues. Universal cross-domain recommendation (UniCDR) further learns a single model to unify multi-domain preferences and support both dual-domain and multi-domain scenarios. However, incorporating source-domain information can even hurt target-domain performance, a phenomenon known as negative transfer. This is because cross-domain signals are not uniformly beneficial: representation learning may capture shortcut cues that break under domain shift, and static fusion may further propagate such unreliable components into the target representation. To explicitly address what to transfer, we propose the Refined Transfer Framework (RTF), an end-to-end architecture that decouples robust representation learning from selective cross-domain fusion. At the representation layer, RTF employs a temperature-scaled attention encoder over the source history and the global history, together with popularity down-weighting to suppress shortcut cues. At the fusion layer, we introduce a domain-routed decoder that aligns multi-source representations in a shared space and selectively fuses auxiliary signals with a global fallback. Finally, we optimize RTF with recommendation supervision and dual contrastive objectives to enhance intra-view robustness and cross-view consistency. Extensive experiments on Amazon benchmarks across four UniCDR scenarios demonstrate that RTF yields competitive and stable improvements over strong baselines.
Large Language Models (LLMs) excel at language understanding but often hallucinate under multi-hop reasoning and schema constraints. Knowledge graphs (KGs), with verifiable symbolic queries, can mitigate this issue. We present RelGraphQA, a training-free, relation-driven framework that anchors LLM generation to KG schema via thresholded FAISS matching and deterministic query synthesis. Natural-language relation mentions are grounded to schema-valid relation types using a lightweight FAISS IndexIVFFlat, retaining only high-similarity candidates. We also early infer head/answer types (and common operators) as structural guards for SPARQL/Cypher. A compact five-stages, four LLM calls pipeline performs (i) type/operator inference, (ii) relation grounding with thresholded schema matching, (iii) single-query synthesis, and (iv) faithful verbalization strictly from executed results. On MetaQA, WebQSP, CWQ, and GrailQA, RelGraphQA achieves competitive Hits@1 scores and demonstrates robust performance across varying threshold settings. Ablations and sensitivity studies confirm the benefit of schema-indexed relation filtering and structural guards. Our code is publicly available at https://github.com/MarisaMagic/RelGraphQA .
In recent years, an increasing number of studies have explored the use of text-attributed graphs to enhance the question-answering capabilities of Large Language Models (LLMs). However, these studies often struggle to effectively integrate structured knowledge and textual information into LLMs. To this end, this paper proposes a novel method to seamlessly integrate text-attributed graphs with pre-trained LLMs. Specifically, the method includes three core components: a graph encoder for capturing graph structural semantics, a graph-text alignment adapter for aligning the semantic spaces of text and graph representations, and a cross-attention module for dynamically fusing graph information into LLM embeddings. Extensive experiments on tasks of commonsense reasoning, scene graph question answering, and multi-hop knowledge graph question answering demonstrate that the method significantly improves LLM’s comprehension of text-attributed graph information, outperforming existing baselines.
Log parsing, the process of converting raw log messages into structured formats, plays a crucial role in automated log analysis. The recent method which first groups logs and then leverages LLMs to extract log templates, has shown promising performance. However, most existing grouping methods rely primarily on statistical features of log data while overlooking the inherent semantic characteristics of log statements, often resulting in inaccurate grouping. Moreover, although LLMs have demonstrated advantages in log parsing, they are typically guided to superficially aggregate grouped logs to produce seemingly correct templates, rather than fully exploiting their capability to verify whether logs within a group truly share the same template. To address these limitations, we propose VocabLog, a vocabulary-driven and LLM-augmented framework for high-performance log parsing. Our approach employs a two-stage semantic grouping strategy to achieve more accurate log grouping, followed by an analysis-then-execution prompt strategy that guides LLMs to analyze grouped logs before performing precise template extraction. Furthermore, an entropy-based extractor and a template merge function are incorporated to further enhance both accuracy and efficiency. Experimental results demonstrate that VocabLog significantly improves both grouping accuracy and template extraction quality, achieving a GA of 99.4
The rapid proliferation of misinformation on social media platforms has rendered fact-checking a pivotal task for natural language processing (NLP). However, current approaches struggle with claim vagueness, particularly in effectively surfacing latent information and capturing the intricate relations in claims. Furthermore, these methods overlook the inherent challenges in evidence reliability: evidence from open-domain sources can suffer from informational incompleteness when presented out-of-context, and it often contains redundant information that is useless or even harmful. To tackle these limitations, we propose an ARgumentation-GUided Evidence-aware framework for fact-checking, namely ARGUE in short. ARGUE consists of two main modules: adaptive prompting based on argumentative reasoning identifies latent information and resolves complex relations to enhance the verifiability of claims, and multi-channel noise-robust evidence-aware network improves evidence quality and reliability by integrating auxiliary information and filtering redundant information. Experiment results on two public fact-checking datasets show superior or comparable performance of ARGUE.
We present DESMA, a distributed discrete-event simulation framework for synchronous multi-asset limit order books. DESMA maps each asset to a simulation kernel for parallel execution and supports cross-asset agents that observe and trade multiple books under a unified discrete-event semantics. To capture cross-asset investment behavior, we introduce a cross-asset proxy agent based on Smooth Prospect Theory (SPT) as a reusable model for generic cross-asset decision making. DESMA combines MPI Remote Memory Access with lock-free single-producer/single-consumer queues for high-throughput inter-process messaging and uses the Chandy–Misra–Bryant conservative protocol to ensure causality without rollback. To reduce synchronization and communication overhead under dense cross-asset interactions, DESMA performs epoch-based reconfiguration by building a kernel-agent dependency hypergraph, applying topology-aware hypergraph partitioning, and jointly updating the communication topology and agent placement. Experiments validate DESMA as a practical platform for multi-asset market simulation by reproducing two classic stylized facts (leptokurtic returns and volatility clustering), demonstrating scalable execution, and showing that hypergraph-based reconfiguration improves timeliness under drifting cross-asset interaction patterns.
To address the critical challenge of balancing energy efficiency with personalized occupant comfort in dynamic building environments, this paper proposes a novel Neuro-Symbolic Control Framework based on Large Language Models (LLMs). The control problem is reformulated as a multi-agent non-cooperative game, introducing persona-driven agents with distinct thermal preferences that negotiate real-time decisions through a strategic counter-bidding mechanism. A core innovation is the Dual-Layer Evolutionary Strategy, which uniquely combines numerical weight optimization with natural language directive generation, enabling a closed-loop control system that evolves from parameter tuning to cognitive strategic reflection. Co-simulation results based on EnergyPlus and data-driven virtual sensors demonstrate that the system effectively adapts to Time-of-Use (TOU) pricing. It significantly reduces operational costs while preventing extreme thermal discomfort, achieving Pareto optimality between energy consumption and comfort, and offering superior interpretability compared to traditional “black-box” models.
With the widespread adoption of network encryption protocols, efficient traffic classification has become a cornerstone of network management and cybersecurity. However, most Transformer-based approaches face a dual challenge: the efficiency bottleneck caused by quadratic computational complexity, and weak robustness in complex and dynamic network scenarios. To address these challenges, this paper presents PFMamba, a lightweight multimodal framework that fuses packet headers and engineered flow statistical features. First, we design a Mamba-based backbone with a dual tokenization strategy that strictly adheres to a flow-based splitting strategy to ensure robust feature extraction free from data leakage. Second, to deeply integrate microscopic packet headers with macroscopic flow statistical features, we introduce a novel Dynamic Adaptive Gated Cross-Attention (DA-GCA) module. Finally, a synergistic pre-training strategy incorporating Masked Autoencoders and Denoising Diffusion (MAE-Diffusion) is designed to enhance protocol semantic understanding. Experimental results on the TLS-120 and Tor-8HS datasets demonstrate that PFMamba outperforms state-of-the-art baselines in both frozen and unfrozen fine-tuning settings. Notably, compared to Pcap-Encoder, our PFMamba-Nano achieves a parameter reduction of over 90
Multi-Instance Multi-Label Learning (MIML) has demonstrated unique advantages in tasks such as image classification and text annotation. However, existing methods still have shortcomings in modeling complex dependencies between samples and correlations among labels. This paper proposes a label correlation embedding based graph attention neural network for multi-instance multi-label learning (LCGAT). Firstly, instances in each bag are constructed into a graph structure, and Graph Attention Network (GAT) is used to model the structural relationships between instances, with edge features introduced to characterize the similarity between instances. Secondly, a correlation-aware module is employed in the label space to capture high-order dependencies among labels, thereby improving the accuracy of label prediction. Furthermore, we introduce a contrastive learning strategy to establish discriminative constraints between instance representations and label representations, enhancing the generalization ability of the model. Experiments on multiple public datasets verify the effectiveness of the proposed method. Compared with mainstream baseline methods, it achieves significant improvements in prediction accuracy and macro-averaged metrics.
Process parameter optimization in optical cable sheath manufacturing is challenging due to highly nonlinear dynamics, strong parameter coupling, and expensive real-world evaluations. We propose a Surrogate Co-evolutionary Incremental Optimization (SCE-IO) framework that combines a task-specific multi-task surrogate model, namely the Optical Cable Sheath Objective Model (OCS-OM), with a Correlation-Guided NSGA-II (CG-NSGA-II) optimizer. To improve robustness under process drift, the framework further incorporates incremental updates using real production feedback. Experiments on real industrial datasets show that SCE-IO consistently outperforms representative baselines, achieving improvements of +38.61
In online education, course recommendation plays an important role in supporting sustained and goal-oriented learning processes. However, noisy non–learning-oriented behaviors and sparse, delayed feedback make it difficult to capture reliable signals that reflect long-term learning progression.To address this issue, we propose Edu-LTER, a reinforcement learning–based course recommendation model that incorporates long-term value signals. Edu-LTER integrates three components: a noise-robust state encoding module, a long-term value enhancement module, and a fusion-based ranking module, enabling unified modeling of behavior representation, value estimation, and decision making. Experimental results on the MOOCCube dataset show that Edu-LTER outperforms strong baselines, achieving an NDCG@20 of 0.3378 with a 7.8
Nowadays, recommendation models based on their approaches to modeling user-item data, exhibit significant limitations in handling complex user behavior patterns and cannot be adaptively applied to multiple scenarios. To address the current algorithms’ insufficient mining of auxiliary interaction information, this paper proposes Instruction-Finetuned Heterogeneous-Aware Graph Neural Network for Temporally Weighted Recommendation Model (IFHAGrec). It models user semantic features through instruction fine-tuning, integrates the modeling of item features, and enhances recommendation performance via context learning, few-shot, and zero-shot learning. Then it constructs a Unified Sentence Vector (USV) processed through a bidirectional GRU to identify item relationships within the heterogeneous graph. Subsequently, after encoding and decoding, it generates candidate sets for multiple item-related tasks under the instruction fine-tuned large model. This approach can capture users’ dynamic interest changes during interactions, thereby more accurately modeling the representations of items within the heterogeneous graph. Extensive experiments on three real-world datasets demonstrate that the IFHAGrec model significantly outperforms state-of-the-art baseline recommendation methods.
With the proliferation of blended learning models, the accurate identification of academically at-risk students is pivotal for implementing personalized educational interventions and ensuring instructional quality. However, real-world educational data presents a dual challenge: extreme class imbalance (with failure rates below 10
Early detection of depression via social media texts is crucial for timely intervention. However, detecting such expressions is challenging due to their subtle, multi-granular, and unevenly distributed nature. While pre-trained models show promise, their high computational cost hinders deployment in resource-limited clinical settings. To address this, we propose MEMBERT, a lightweight inductive-bias-driven framework. It integrates MobileBERT with Mamba-2 modules for long-range emotional evolution modeling and an ECA-style gating mechanism for emotion-sensitive feature fusion, enhancing fine-grained modeling with minimal overhead. Experiments show MEMBERT achieves a Macro-F1 of 0.7666 and a False Negative Rate of 0.2068, outperforming its lightweight backbone and achieving competitive performance with BERT under efficiency constraints while increasing parameters by only 23.1
