Despite their impressive capabilities, large language models (LLMs) frequently generate hallucinations. Previous work shows that their internal states encode rich signals of truthfulness, yet the origins and mechanisms of these signals remain unclear. In this paper, we demonstrate that truthfulness cues arise from two distinct information pathways: (1) a Question-Anchored pathway that depends on question–answer information flow, and (2) an Answer-Anchored pathway that derives self-contained evidence from the generated answer itself. First, we validate and disentangle these pathways through attention knockout and token patching. Afterwards, we uncover notable and intriguing properties of these two mechanisms. Further experiments reveal that (1) the two mechanisms are closely associated with LLM knowledge boundaries; and (2) internal representations are aware of their distinctions. Finally, building on these insightful findings, two applications are proposed to enhance hallucination detection performance. Overall, our work provides new insight into how LLMs internally encode truthfulness, offering directions for more reliable and self-aware generative systems.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Learning to Think on Hypergraph: HyperCoT for Structure-Guided N-ary Knowledge Graph Completion
PDF ↗N-ary knowledge graph completion (KGC) aims to infer missing components in facts with multiple entities under distinct semantic roles, commonly formulated as a knowledge hypergraph link prediction task. Most embedding-based approaches score individual hyperedges relying on enriched structural representations, but overlook intermediate propagation states containing complementary local and global structural evidence. Despite their capability to generate chain-of-thought (CoT) representations for the classical KGC task, large language models (LLMs) struggle with hypergraph structure involving multiple facts, while current hypergraph QA methods only provide LLMs with a single query signal rather than path-level evidence. These limitations hinder the transferability of existing methods, especially those leveraging LLMs, to solve the knowledge hypergraph link prediction problem. To bridge this gap, we propose HyperCoT, a structure-aware approach that models multi-hop structural reasoning as a depth-sensitive progressive evidence accumulation process. It constructs a Graphical Chain-of-Thought (Graph-CoT) by aggregating role-aware hyperedge states along strongly correlated reasoning paths, and injects the resulting path-level structural evidence into each token in query and candidate entities to prompt LLMs. Experiments on three real-world datasets demonstrate that HyperCoT consistently outperforms strong n-ary KGC baselines, particularly in high arity and structural sparsity scenarios, meanwhile yielding interpretable multi-hop reasoning traces.
Revisiting Metric Reliability for Fine-grained Evaluation of Machine Translation and Summarization in Indian Languages
PDF ↗While automatic metrics drive progress in Machine Translation (MT) and Text Summarization (TS), existing metrics have been developed and validated almost exclusively for English and other high-resource languages. This narrow focus leaves Indian languages—spoken by over 1.5 billion people—largely overlooked, casting doubt on the universality of current evaluation practices. To address this gap, we introduce ITEM, a large-scale benchmark that systematically evaluates the alignment of 26 automatic metrics with human judgments across six major Indian languages, enriched with fine-grained annotations. Our extensive evaluation—covering agreement with human judgments, sensitivity to outliers, language-specific reliability, inter-metric correlations, and resilience to controlled perturbations—reveals four central findings: (1) LLM-based evaluators show the strongest alignment with human judgments at both segment and system levels; (2) outliers exert a significant impact on metric-human agreement; (3) In TS, metrics are more effective at capturing content fidelity, whereas in MT, they better reflect fluency; and (4) Metrics differ in their robustness and sensitivity when subjected to diverse perturbations. Collectively, these findings offer critical guidance for advancing metric design and evaluation in Indian languages.
Cognitive Scaffold: From Fluid Context to Crystallized Memory for Long-Horizon DeepResearch Agents
PDF ↗Scaling LLM-based agents to long-horizon deep research is constrained by the context-noise trade-off, where linear history accumulation degrades reasoning and dilutes fine-grained evidence. To address this, we introduce the Cognitive Scaffold, a factorized memory architecture that decouples the cognitive state into a Fluid Working Context for immediate reasoning and a persistent Knowledge Graph for long-term retention. Unlike unstructured summarization, our framework employs a Rejection Sampling Fine-Tuning (RFT) pipeline to crystallize saturated context into structured event snapshots, strictly enforcing atomic constraints to preserve numerical values and entities. During reasoning, a thought-driven dual-path retrieval mechanism enables the agent to proactively recover precise evidence. Empirical evaluations on Xbench-DeepSearch, BrowseComp-ZH, and GAIA demonstrate that Cognitive Scaffold consistently outperforms baselines, achieving 74.7% Avg@3 and 87.0% Pass@3 on Xbench-DeepSearch, 48.5% Avg@3 and 65.9% Pass@3 on BrowseComp-ZH, and 72.8% Avg@3 and 88.3% Pass@3 on GAIA, while reducing compression hallucinations to 5.3%. We open-source our codebase to facilitate future research.
Integrating textual graphs into Large Language Models (LLMs) is promising for complex graph-based QA. However, a key bottleneck is retrieving informative yet compact subgraphs that fit the LLM context. Existing retrievers often struggle, relying either on shallow embedding similarity or costly interactive policies that require excessive supervision.To address these challenges, we introduce Graph-S^3, an agentic textual graph reasoning framework featuring an LLM-based retriever trained with synthetic stepwise supervision. Rather than relying on final answer rewards—which often yield sparse and unstable signals—we optimize the retriever by evaluating each step against offline-extracted golden subgraphs.Our approach distills golden subgraphs via a specialized data synthesis pipeline to formulate dense rewards, facilitating a two-stage training scheme that effectively learns the interactive graph exploration policy.Based on extensive experiments on three common datasets in comparison with seven strong baselines, our approach achieves an average improvement of 15.6% in accuracy and 17.2% in F_1 score. The advantage is even higher in more complicated multi-hop reasoning tasks.
Current defenses for Large Language Models (LLMs) often suffer from a ”memory gap”: parameter-modifying methods are computationally rigid, while inference-time filters cannot retain or reuse defense knowledge across interactions. To address this, we propose SafetyMem, a novel framework that secures LLMs through a dual-component safety memory system. SafetyMem consists of Semantic Safety Memory (SSM), which consolidates diverse jailbreak attempts into a structured knowledge base of attack patterns, and Episodic Safety Memory (ESM), which maintains an evolving set of procedural rules refined from historical detection failures. Unlike static defenses, SafetyMem allows the model to ”remember” and adapt to emerging adversarial strategies without parameter retraining. To further enhance robustness, we introduce an adversarial memory expansion mechanism that proactively generates challenging variants to solidify these memories. Experiments on standard and stealthy jailbreak benchmarks show that SafetyMem substantially reduces attack success rates while preserving efficiency and interpretability, consistently outperforming state-of-the-art baselines across multiple LLMs.
While Large Language Models (LLMs) excel at algorithmic code generation, they struggle with front-end development, where correctness is judged on rendered pixels and interaction. We present ReLook, an agentic, vision-grounded reinforcement learning framework that empowers an agent to close a robust generate–diagnose–refine loop by invoking a multimodal LLM (MLLM) as a tool. During training, the agent employs an MLLM-in-the-loop to serve as a visual critic, evaluating code via screenshots and providing actionable feedback. Crucially, we enforce a strict zero-reward policy for invalid renders to guarantee renderability and mitigate reward hacking. To prevent behavioral collapse, we introduce Forced Optimization, a strict acceptance rule that admits only improving revisions, yielding monotonically better trajectories. At inference, we decouple the critic and run a lightweight, critic-free self-edit cycle, keeping latency comparable to base decoding while retaining most of the gains. Across three widely used benchmarks, ReLook consistently outperforms strong baselines in vision-grounded front-end code generation, highlighting the benefits of agentic perception, visual rewards, and training–inference decoupling.
AIRCoder: Adaptive Integration of Multi-dimensional Retrieval for Repository-level Code Completion
PDF ↗Repository-level code completion relies on retrieval strategies to select relevant context from large codebases. Most existing methods employ single-dimensional retrieval based on textual similarity or dependency existence, leading to inconsistent performance across completion scenarios. Even task-adaptive and hybrid methods incur high computational costs while failing to explicitly consider structural hierarchy for retrieving functionally similar code. We introduce **AIRCoder**, a multi-dimensional retrieval framework that combines eight complementary metrics across three dimensions: textual similarity, dependency existence, and structural hierarchy. By proposing a structure-preserving chunking strategy and lightweight fusion module, AIRCoder learns context-dependent weights to adaptively integrate retrieval metrics for each query. Experiments on CrossCodeEval and RepoEval demonstrate that AIRCoder achieves an average improvement of **4.63%** in exact match over the best baseline, with **10.2\times** higher efficiency and strong cross-language generalization across Python, Java, C#, and TypeScript.
Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for improving reasoning capabilities. However, training RLVR with Mixture-of-Experts (MoE) policies remains fragile and is often prone to reward collapse.We identify a MoE-specific source of instability, referred to as router shift (RS), where changes in expert routing across policy updates exacerbate off-policy mismatch. This effect leads to increasingly volatile importance-ratio signals and bursty clipping behavior, which consistently precede training collapse.Motivated by this diagnosis, we propose Router-Shift Policy Optimization (RSPO). RSPO computes a per-token router-shift ratio conditioned on the previously activated experts, applies stop-gradient and a lower-bound floor, and softly rescales importance ratios prior to clipping and aggregation. This design explicitly accounts for routing-induced distributional drift during off-policy optimization.We evaluate the effect of RSPO under two settings: a synthetic countdown task and real-world reasoning tasks on MATH and Code. Across both settings, RSPO achieves better performance and exhibits greater stability compared to recent MoE-based RLVR methods.
Multimodal Retrieval-Augmented Generation (MRAG) enhances reasoning capabilities by integrating external knowledge. However, existing benchmarks primarily focus on simple image-text interactions, overlooking complex visual formats like charts that are prevalent in real-world applications. In this work, we introduce a novel task, \textbf{Chart-based MRAG}, to address this limitation. To generate high-quality evaluation samples, we propose \textbf{CHARGE} (\textbf{CHAR}t-based document question-answering \textbf{GE}neration), a semi-automatic framework for generating evaluation samples through multi-modal keypoint extraction, knowledge graph construction, and qa pair synthesis.By combining CHARGE with expert validation, we construct \textbf{Chart-MRAG Bench}, a comprehensive benchmark for chart-based MRAG evaluation, featuring 4,738 question-answering pairs across 8 domains from real-world documents.Our experiments reveal three critical limitations in current approaches: (1) unified multimodal embedding retrieval methods struggles in chart-based scenarios, (2) even with ground-truth retrieval, state-of-the-art Multimodal Large Language Models (MLLMs) achieve only 71.15% Correctness and 80.74% Coverage scores, and (3) Widely-used MLLMs demonstrate consistent text-over-visual modality bias. These findings highlight great challenges in processing information-dense visual formats. We will make our code and dataset publicly available.
Tables contain rich structured information, yet when stored as images their contents remain "locked" within pixels. Converting table images into LaTeX code enables faithful digitization and reuse, but current multimodal large language models (MLLMs) often fail to preserve structural, style, or content fidelity. Conventional post-training with reinforcement learning (RL) typically relies on a single aggregated reward, leading to reward ambiguity that conflates multiple behavioral aspects and hinders effective optimization. We propose Component-Specific Policy Optimization (CSPO), an RL framework that disentangles optimization across LaTeX tables components—structure, style, and content. In particular, CSPO assigns component-specific rewards and backpropagates each signal only through the tokens relevant to its component, alleviating reward ambiguity and enabling targeted component-wise optimization. To comprehensively assess performance, we introduce a set of hierarchical evaluation metrics. Extensive experiments demonstrate the effectiveness of CSPO, underscoring the importance of component-specific optimization for reliable structured generation.
Multimodal Large Language Models (MLLMs) have been increasingly used as automatic evaluators—a paradigm known as *MLLM-as-a-Judge*. However, their reliability and vulnerabilities to biases remain underexplored. We find that many MLLM judges fail to reliably integrate key visual or textual cues, yielding unreliable evaluations when evidence is missing or mismatched, and exhibiting instability under semantically irrelevant perturbations. To address this, we systematically define *Compositional Bias* in MLLM-as-a-Judge systems and introduce **MM-JudgeBias**, a benchmark for evaluating it. MM-JudgeBias introduces controlled perturbations across Query, Image, and Response, and evaluates model behavior via two complementary metrics: *Bias-Deviation (BD)* for sensitivity and *Bias-Conformity (BC)* for stability. Our dataset of over 1,800 curated and refined multimodal samples, drawn from 29 source benchmarks, enables a fine-grained diagnosis of nine bias types across diverse tasks and domains. Experiments on 26 state-of-the-art MLLMs reveal systematic modality neglect and asymmetric evaluation tendencies, underscoring the need for more reliable judges.
Large Language Models (LLMs) increasingly rely on agentic capabilities—iterative retrieval, tool use, and decision-making—to overcome the limits of static, parametric knowledge. Yet existing agentic frameworks treat external information as unstructured text and fail to leverage the topological dependencies inherent in real-world data. To bridge this gap, we introduce Agentic Graph Learning (AGL), a paradigm that reframes graph learning as an interleaved process of topology-aware navigation and LLM-based inference. Specifically, we propose AgentGL, the first reinforcement learning (RL)–driven framework for AGL. AgentGL equips an LLM agent with graph-native tools for multi-scale exploration, regulates tool usage via search-constrained thinking to balance accuracy and efficiency, and employs a graph-conditioned curriculum RL strategy to stabilize long-horizon policy learning without step-wise supervision. Across diverse Text-Attributed Graph (TAG) benchmarks and multiple LLM backbones, AgentGL substantially outperforms strong GraphLLMs and GraphRAG baselines, achieving absolute improvements of up to 17.5% in node classification and 28.4% in link prediction. These results demonstrate that AGL is a promising frontier for enabling LLMs to autonomously navigate and reason over complex relational environments. The code is publicly available at https://github.com/sunyuanfu/AgentGL.
Low-Rank Adaptation (LoRA) has emerged as a prominent solution to mitigate the communication and computation costs in federated fine-tuning of Large Language Models (LLMs). However, we observe that even within low-rank adapters, a substantial portion of parameters manifest negligible updates during federated training, leading to redundant communication and wasted local computation. To address this, we propose GMFL, a plug-and-play layer freezing mechanism designed to seamlessly integrate with existing federated fine-tuning frameworks. Specifically, the server monitors the global update magnitude of each LoRA layer to dynamically generate freezing masks. These masks are updated periodically with a fixed freezing rate, ensuring stable convergence by robustly identifying “saturated” layers. Theoretical analysis confirms the convergence of GMFL, where the freezing mechanism yields a bounded error that scales with client heterogeneity. Extensive experiments across multiple tasks (GLUE, Commonsense Reasoning, Math Reasoning and General Generation) demonstrate that GMFL reduces communication overhead and lowers computational costs while preserving the performance of the underlying federated fine-tuning methods. Our work provides a practical, versatile solution for deploying large-scale federated LLM fine-tuning in resource-constrained environments. Our code is available at: https://github.com/tunx-cyber/GMFL.
Detecting hallucinations in Retrieval-Augmented Generation remains a challenge. Prior approaches attribute hallucinations to a binary conflict between internal knowledge stored in FFNs and the retrieved context. However, this perspective is incomplete, failing to account for the impact of other components of the LLM, such as the user query, previously generated tokens, the self token, and the final LayerNorm adjustment. To comprehensively capture the impact of these components on hallucination detection, we propose TPA which mathematically attributes each token’s probability to seven distinct sources: Query, RAG Context, Past Token, Self Token, FFN, Final LayerNorm, and Initial Embedding. This attribution quantifies how each source contributes to the generation of the next token. Specifically, we aggregate these attribution scores by Part-of-Speech (POS) tags to quantify the contribution of each model component to the generation of specific linguistic categories within a response. By leveraging these patterns, such as detecting anomalies where Nouns rely heavily on LayerNorm, TPA effectively identifies hallucinated responses. Extensive experiments show that TPA achieves state-of-the-art performance.
Code Language Models (CodeLLMs) traditionally learn attention based solely on statistical input-output token correlations ("machine attention"). In contrast, human developers rely on intuition, selectively fixating on semantically salient tokens during program comprehension. We present EyeMulator, a model-agnostic technique to align CodeLLM attention with human visual attention without architectural changes. By extracting scan paths from eye-tracking data, we derive token-level attention weights used to augment the loss function during fine-tuning. This induces the model to mimic human focus. Our evaluation across StarCoder, Llama-3.2, and DeepSeek-Coder shows that EyeMulator significantly outperforms baselines, achieving gains of over 30 CodeBLEU points in translation and up to 22 BERTScore points in summarization. Ablation studies confirm that these gains stem directly from replicating human attention dynamics. Artifacts are available at https://zenodo.org/records/17205682.
WebSynthesis: World Model-Guided Monte Carlo Tree Search for Efficient WebAgent Trajectory Synthesis
PDF ↗Recent advances in large language models (LLMs) have enabled increasingly capable web agents, yet training such agents still relies on high-quality interaction trajectories that are difficult to obtain at scale. We identify two key challenges: (1) Infrastructure Overhead, where network instability and website access restrictions limit data collection scalability; and (2) Constrained Exploration, where irreversible state transitions preclude tree-based search and thus limit trajectory diversity. To address these challenges, we introduce WebSynthesis, a framework for scalable trajectory synthesis. WebSynthesis employs an LLM-based World Model to simulate state transitions without network dependencies, and integrates Monte Carlo Tree Search to enable reversible exploration over the simulated state space. Experiments on WebArena, WebVoyager, and Mind2Web-Online demonstrate that agents trained exclusively on synthesized trajectories outperform those trained on real-world data, providing a viable alternative to costly real-world data collection.
Retrieval-Augmented Generation (RAG) has long been a promising paradigm for enhancing large language models (LLMs) with external knowledge. Traditional embedding-based methods for graph construction can capture semantic similarity but struggle to establish fine-grained, interpretable logical relationships. Recently, Graph-enhanced RAG (GraphRAG) has gained increasing popularity for its capability in modeling logical relationships. However, graph construction requires extensive token consumption for triple extraction and summarization, making it costly and slow. Accordingly, we propose MeshRAG, a novel framework for mining efficient structures via hashing to enhance RAG. We adopt an inductive paradigm in which global graph structure emerges from local hash collisions rather than explicit symbolic extraction. By replacing neural embedding search with lightweight and bitwise operations, MeshRAG automates a simple and rapid graph construction process. Furthermore, the hash collision mechanism provides transparent evidence for logical connections and retrieval decisions. Experimental results show that MeshRAG outperforms existing baselines, while its graph construction requires no GPU resources or token budget and can structure over ten thousand chunks in a few minutes.
Bargaining is often regarded as a logical arena rather than an art or a matter of intuition, yet Large Language Models (LLMs) still struggle to navigate it due to limited strategic depth and difficulty adapting to complex human factors. Current benchmarks rarely capture this limitation. To bridge this gap, we present a utility feedback centric framework. Our contributions are: (i) AgoraBench, a new benchmark spanning nine challenging settings (e.g., deception, monopoly) that supports diverse strategy modeling; (ii) human-aligned, economically grounded metrics derived from utility theory. This is operationalized via agent utility, negotiation power, and acquisition ratio that implicitly measure how well the negotiation aligns with human preference and (iii) a human preference grounded dataset with learning pipeline that strengthens LLMs’ bargaining ability through both prompting and finetuning. Empirical results indicate that baseline LLM strategies often diverge from human preferences, while our mechanism substantially improves negotiation performance, yielding deeper strategic behavior and stronger opponent awareness.
Large Language Models have emerged as powerful tools for automating Register-Transfer Level (RTL) code generation, yet they face critical limitations: existing approaches typically fail to simultaneously optimize functional correctness and hardware efficiency metrics such as Power, Performance, and Area (PPA). Methods relying on supervised fine-tuning commonly produce functionally correct but suboptimal designs due to the lack of inherent mechanisms for learning hardware optimization principles. Conversely, external post-processing techniques aiming to refine PPA performance after generation often suffer from inefficiency and do not improve the LLMs’ intrinsic capabilities.To overcome these challenges, we propose ChipSeek, a novel hierarchical reward based reinforcement learning framework designed to encourage LLMs to generate RTL code that is both functionally correct and optimized for PPA metrics. Our approach integrates direct feedback from EDA simulators and synthesis tools into a hierarchical reward mechanism, facilitating a nuanced understanding of hardware design trade-offs. Through Curriculum-Guided Dynamic Policy Optimization (CDPO), ChipSeek enhances the LLM’s ability to generate high-quality, optimized RTL code. Evaluations on standard benchmarks demonstrate ChipSeek’s superior performance, achieving state-of-the-art functional correctness and PPA performance. Furthermore, it excels in specific optimization tasks, consistently yielding highly efficient designs when individually targeting fine-grained optimization goals such as power, delay, and area. The artifact is open-source in https://github.com/rong-hash/chipseek.