Evaluating whether large language models (LLMs) capture the structureof natural language beyond local fluency remains an open challenge.Existing evaluation methods, largely based on task performance orshort-context behavior, provide limited insight into the long-rangestatistical organization of generated text.We propose a complementary evaluation framework based on repeatedsubsequences. By analyzing their distribution across scales andrelating it to higher-order Rényi entropies, we probe how textsreuse previously established structure under finite-lengthconditions. Experiments on human-written texts and length-matchedGPT-generated texts show that,while power-law models can describerestricted ranges of block length, the observed entropy growth isoften equally or better characterized by logarithmic–power forms.Across datasets, natural language exhibits stable entropy-growthpatterns over accessible ranges, with consistent average behavior despite variability across individual texts. In contrast,GPT-generated texts show systematic and statistically significantshifts in estimated exponents with model size.These results demonstrate that repeated-subsequence entropyprovides a quantitative structural diagnostic that revealssystematic differences in long-range organization,distinguishing natural language from state-of-the-art LLM outputsbeyond surface-level fluency.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Dual Activation-Weight Sparsity: A Training-Free Framework for Efficient Large Language Model Compression
PDF ↗Large language models (LLMs) excel at natural language tasks but face deployment challenges due to computational demands. We introduce Dual Activation-Weight Sparsity (DAWS), a training-free framework that jointly exploits activation and weight sparsity through magnitude-based routing. Systematic analysis of pretrained transformers reveals two key observations: (1) the activation energy is concentrated in a few neurons, and (2) activation and weight sparsity patterns are complementary between attention and FFN layers. DAWS employs a three-tier routing strategy: high-magnitude activations pass through full-precision weights to preserve critical pathways, medium-magnitude activations use magnitude-pruned sparse weights for efficiency, and low-magnitude activations are directly discarded. Unlike prior work that uses activation-aware pruning methods like WANDA, our approach uses direct magnitude-based pruning, which we show is more robust to sample-level variations. Experiments on Llama and Mistral models demonstrate that DAWS maintains >98% of dense model performance at 50% sparsity, outperforming WANDA, TEAL, and R-Sparse.
Visualizations play a pivotal role in daily communication in an increasingly data-driven world. Research on multimodal large language models (MLLMs) for automated chart understanding has accelerated massively, with steady improvements on standard benchmarks. However, for MLLMs to be reliable, they must be robust to misleading visualizations, i.e., charts that distort the underlying data, leading readers to draw inaccurate conclusions. Here, we uncover an important vulnerability: MLLM question-answering (QA) accuracy on misleading visualizations drops on average to the level of the random baseline. To address this, we provide the first comparison of six inference-time methods to improve QA performance on misleading visualizations, without compromising accuracy on non-misleading ones. We find that two methods, table-based QA and redrawing the visualization, are effective, with improvements of up to 19.6 percentage points. We make our code and data available.
Enhancing the reasoning capabilities of Large Language Models (LLMs) is a key strategy for building agents that ”think then act”. However, recent observations, like OpenAI’s o3, suggest a paradox: stronger reasoning often coincides with increased hallucination, yet no prior work has systematically examined whether reasoning enhancement itself causes tool hallucination. We address this gap with the central question: Does strengthening reasoning increase tool hallucination? To answer this, we introduce SimpleToolHalluBench, a diagnostic benchmark measuring tool hallucination in two failure modes: (i) no tool available, and (ii) only distractor tools available. Through controlled experiments, we establish three key findings. First, we demonstrate a causal relationship: progressively enhancing reasoning through RL increases tool hallucination proportionally with task performance gains. Second, this effect transcends overfitting—training on non-tool tasks (e.g., mathematics) still amplifies subsequent tool hallucination. Third, the effect is method-agnostic, appearing when reasoning is instilled via supervised fine-tuning and when it is merely elicited at inference by switching from direct answers to step-by-step thinking. We also evaluate mitigation strategies including Prompt Engineering and Direct Preference Optimization (DPO), revealing a fundamental reliability–capability trade-off: reducing hallucination consistently degrades utility. Mechanistically, Reasoning RL disproportionately collapses tool-reliability–related representations, and hallucinations surface as amplified divergences concentrated in late-layer residual streams. These findings reveal that current reasoning enhancement methods inherently amplify tool hallucination, highlighting the need for new training objectives that jointly optimize for capability and reliability. Our implementation is provided at https://github.com/albert-y1n/Reasoning_Trap.
RSDA: Restoring Stale Data Affinity via Dynamic Renovation Strategy for Mitigating Data Scarcity
PDF ↗High-quality data is the cornerstone of advancing large language models. However, the field currently faces a critical dilemma: the supply of premium data is nearing depletion, while vast stale corpora remain underutilized. Our empirical analysis reveals that training models on such data directly often leads to performance degradation. We attribute this phenomenon to the data affinity gap, a misalignment stemming from the model’s inability to effectively comprehend the data or inherent quality defects. To bridge this gap, we propose Restoring Stale Data Affinity (RSDA) framework. First, utilizing our proposed potential entropy metric, RSDA quantifies the latent value of samples to effectively identify stale data with higher renovation potential. Subsequently, the framework employs a dynamic renovation strategy selection mechanism to determine the optimal component-level strategy for each instance, transforming low-affinity stale samples into high-quality training data. Comprehensive experimental results demonstrate that RSDA effectively enhances data affinity, achieving performance improvements using less than 10% of the data volume, thereby underscoring that the latent potential of stale corpora remains largely untapped. The code is available at https://github.com/wenfiii/RSDA.
The advent of Large Language Models (LLMs) has significantly revolutionized web search. The emergence of LLM-based Search Agents marks a pivotal shift towards deeper, dynamic, autonomous information seeking. These agents can comprehend user intentions and environment context and execute multi-turn retrieval with dynamic planning, extending search capabilities far beyond the web. Leading examples like OpenAI’s Deep Research highlight their potential for deep information mining and real-world applications. This survey provides the first systematic analysis of search agents. We comprehensively analyze and categorize existing works from the perspectives of architecture, optimization, application, and evaluation, ultimately identifying critical open challenges and outlining promising future research directions in this rapidly evolving field.
Recent advancements in LLM-based multi-agent systems have demonstrated remarkable collaborative capabilities across complex tasks. To enhance the overall efficiency, existing methods often rely on aggressive graph topology evolution for agents (e.g., node or edge pruning), which risks prematurely discarding valuable agents due to transient issues such as hallucinations or temporary knowledge gaps. However, such hard pruning overlooks the potential for "zombie" agents to recover and contribute in subsequent discussion rounds. In this paper, we propose AgentRevive, a Markov state-aware framework for resilient multi-agent evolution. Our approach dynamically manages agent collaboration through soft state transitions, implemented via two key components: (1) State-Aware Policy Learning: Agent states are divided into "Active", "Standby", and "Terminated", selectively propagating messages based on agent memory. The policy employs a risk estimator to optimize agent state transitions by assessing hallucination risk, minimizing the influence of unreliable nodes while safeguarding valuable ones. (2) State-Aware Edge Optimization: Subgraph edges are pruned according to states learned from the policy, permanently removing "Terminated" nodes and retaining "Standby" nodes for subsequent rounds to observe potential future contributions. Extensive experiments on general reasoning, domain-specific, and hallucination challenge tasks show that our method consistently outperforms strong baselines and significantly reduces token consumption through state-aware agent scheduling.
Emotion is a core paralinguistic feature in voice interaction. It is widely believed that emotion understanding models learn fundamental representations that transfer to synthesized speech, making emotion understanding results a plausible reward or evaluation metric for assessing emotional expressiveness in speech synthesis. In this work, we critically examine this assumption by systematically evaluating Speech Emotion Recognition (SER) on synthesized speech across datasets, discriminative and generative SER models, and diverse synthesis models. We find that current SER models can not generalize to synthesized speech, largely because speech token prediction during synthesis induces a representation mismatch between synthesized and human speech. Moreover, generative Speech Language Models (SLMs) tend to infer emotion from textual semantics while ignoring paralinguistic cues. Overall, our findings suggest that existing SER models often exploit non-robust shortcuts rather than capturing fundamental features, and paralinguistic understanding in SLMs remains challenging.
Retrieval-Augmented Generation (RAG) has emerged as a powerful framework for knowledge-intensive tasks, yet its effectiveness in long-context scenarios is often bottlenecked by the retriever’s inability to distinguish sparse yet crucial evidence. Standard retrievers, optimized for query-document similarity, frequently fail to align with the downstream goal of generating a precise answer. To bridge this gap, we propose a novel fine-tuning framework that optimizes the retriever for Answer Alignment. Specifically, we first identify high-quality positive chunks by evaluating their sufficiency to generate the correct answer. We then employ a curriculum-based contrastive learning scheme to fine-tune the retriever. This curriculum leverages LLM-constructed Knowledge Graphs (KGs) to generate augmented queries, which in turn mine progressively challenging hard negatives. This process trains the retriever to distinguish the answer-sufficient positive chunks from these nuanced distractors, enhancing its generalization. Extensive experiments on 10 datasets from the Ultradomain and LongBench benchmarks demonstrate that our fine-tuned retriever achieves state-of-the-art performance, improving 14.5% over the base model without substantial architectural modifications and maintaining strong efficiency for long-context RAG. Our work presents a robust and effective methodology for building truly answer-centric retrievers.
Mem2ActBench: A Benchmark for Evaluating Long-Term Memory Utilization in Task-Oriented Autonomous Agents
PDF ↗Large Language Model (LLM)-based agents are increasingly deployed for complex, tool-based tasks where long-term memory is critical to driving actions. Existing benchmarks, however, primarily test an agent’s ability to passively retrieve isolated facts in response to explicit questions. They fail to evaluate the more crucial capability of actively applying memory to execute tasks. To address this gap, we introduce Mem2ActBench, a benchmark for evaluating whether agents can proactively leverage long-term memory to execute tool-based actions by selecting appropriate tools and grounding their parameters. The benchmark simulates persistent assistant usage, where users mention the same topic across long, interrupted interactions and expect previously established preferences and task states to be implicitly applied. We build the dataset with an automated pipeline that merges heterogeneous sources (ToolACE, BFCL, OASST1), resolves conflicts via consistency modeling, and synthesizes 2,029 sessions with 12 user–assistant–tool turns on average. From these memory chains, a reverse-generation method produces 400 tool-use tasks, with human evaluation confirming 91.3% are strongly memory-dependent. Experiments on seven memory frameworks show that current systems remain inadequate at actively utilizing memory for parameter grounding, highlighting the need for more effective approaches to evaluate and improve memory application in task execution. Code and data are available at https://github.com/Cantaloupe-M/Mem2ActBench.
As video large language models (Video-LLMs) become increasingly integrated into real-world applications that demand grounded multimodal reasoning, ensuring their factual consistency and reliability is of critical importance. However, sycophancy, the tendency of these models to align with user input even when it contradicts the visual evidence, undermines their trustworthiness in such contexts. Current sycophancy research has largely overlooked its specific manifestations in the video-language domain, resulting in a notable absence of systematic benchmarks and targeted evaluations to understand how Video-LLMs respond under misleading user input. To fill this gap, we propose ViSE (Video-LLM Sycophancy Benchmarking and Evaluation), the first benchmark designed to evaluate sycophantic behavior in state-of-the-art Video-LLMs across diverse question formats, prompt biases, and visual reasoning tasks. Specifically, ViSE pioneeringly brings linguistic perspectives on sycophancy into the video domain, enabling fine-grained analysis across multiple sycophancy types and interaction patterns. Furthermore, we propose two potential training-free mitigation strategies revealing potential paths for reducing sycophantic bias: (i) enhancing visual grounding through interpretable key-frame selection and (ii) steering model behavior away from sycophancy via targeted, inference-time intervention on its internal neural representations. Our code is available at https://github.com/William030422/Video-Sycophancy.
Large language models (LLMs) have been applied to sequential decision-making tasks like multi-armed bandits (MAB), where an LLM is tasked with selecting arms in each iteration. However, this direct arm selection approach is often suboptimal. We propose an alternative method combining classical MAB algorithms with LLMs. Specifically, we use a classical MAB framework and leverage the in-context learning capability of LLMs for reward prediction. First, we integrate the LLM-based predictor into Thompson sampling (TS) with a decaying temperature schedule to balance exploration and exploitation. We also incorporate the predictor into a regression oracle-based MAB algorithm with explicit exploration. Additionally, we extend our TS-based algorithm to dueling bandits, where only preference feedback between arm pairs is available, requiring significant algorithmic modifications. Our empirical evaluations on synthetic MAB tasks show that our algorithms outperform LLM-based direct arm selection. In experiments on real-world text datasets, we demonstrate that, in tasks where arms lack exploitable semantic meaning, our approach delivers significantly better performance than direct arm selection.
Multimodal Large Language Models (MLLMs) have become widely deployed, yet their safety alignment remains fragile under adversarial inputs. Previous work has shown that increasing inference steps can disrupt safety mechanisms and lead MLLMs to generate attacker-desired harmful content. However, most existing attacks focus on increasing the complexity of the modified visual task itself and do not explicitly leverage the model’s own reasoning incentives. This leads to them underperforming on reasoning models (Models with Chain-of-Thoughts) compared to non-reasoning ones (Models without Chain-of-Thoughts). If a model can think like a human, can we influence its cognitive-stage decisions so that it proactively completes a jailbreak? To validate this idea, we propose GAMBIT (Gamified Adversarial Multimodal Breakout via Instructional Traps), a novel multimodal jailbreak framework that decomposes and reassembles harmful visual semantics, then constructs a gamified scene that drives the model to explore, reconstruct intent, and answer as part of winning the game. The resulting structured reasoning chain increases task complexity in both vision and text, positioning the model as a participant whose goal pursuit reduces safety attention and induces it to answer the reconstructed malicious query. Extensive experiments on popular reasoning and non-reasoning MLLMs demonstrate that GAMBIT achieves high Attack Success Rates (ASR), reaching 92.13% on Gemini 2.5 Flash, 91.20% on QvQ-MAX, and 85.87% on GPT-4o, significantly outperforming baselines. Warning: This paper contains unsafe and offensive examples.
Agentic learning increasingly hinges on interaction, yet real-world experience is expensive, limited, and often irreversible at inference time. World models promise to mitigate these limitations, but it remains unclear whether large language models can actually serve as reliable world models, and deliver concrete benefits to downstream agents. We investigate these questions in text-based environments, a controlled testbed that reframes language modeling as next-state prediction under interaction. We propose a three-level framework to evaluate LLM-based world models: (i) fidelity and consistency, (ii) scalability and robustness, and (iii) agent utility. Across five representative environments, we show that sufficiently trained world models capture coherent environment dynamics, scale predictably with data and model capacity, and unlock tangible agent improvements—for example, action verification boosts GPT-4o by 5.5% on WebShop, and warm-started RL achieves a 15% gain on SciWorld. Crucially, these benefits hinge on behavioral coverage and environment complexity, sharply characterizing when world modeling meaningfully advances agent learning.
Large Language Models (LLMs) face severe challenges in long-context processing, including quadratic computational costs, information forgetting, and the context fragmentation inherent in Retrieval-Augmented Generation (RAG). We introduce LycheeMemory, a cognitively inspired framework that enables efficient long-context inference via chunk-wise compression and selective memory recall, rather than processing all raw tokens. LycheeMemory segments the input into chunks and encodes each into compressed KV-cache style representations using a Compressor. A Gate then dynamically selects relevant memory blocks, which a Reasoner iteratively processes with an evolving working memory to solve downstream tasks. The Compressor and Reasoner are jointly optimized via end-to-end reinforcement learning, while the Gate is trained separately as a classifier. Experimental results demonstrate that LycheeMemory achieves competitive accuracy (up to 82% in ablation variants) on multi-hop reasoning benchmarks (e.g., RULER-HQA), successfully extrapolates context length from 7K to 1.75M, and provides a favorable accuracy–efficiency trade-off against strong long-context baselines. Notably, compared to MemAgent, LycheeMemory achieves an average 2× reduction in peak GPU memory usage and a 6× speedup during inference.
A fundamental obstacle to causal discovery from text is the lack of causally annotated text data for use as ground truth, due to high annotation costs. This motivates an important task of generating text with causal graph annotations. Early template-based generation methods sacrifice text naturalness in exchange for high causal graph annotation accuracy. Recent Large Language Model (LLM)-dependent methods directly generate natural text from target graphs through LLMs, but do not guarantee causal graph annotation accuracy. Therefore, we propose iTAG, which performs real-world concept assignment to nodes before converting causal graphs into text in existing LLM-dependent methods. iTAG frames this process as an inverse problem with the causal graph as the target, iteratively examining and refining concept selection through Chain-of-Thought (CoT) reasoning so that the induced relations between concepts are as consistent as possible with the target causal relationships described by the causal graph. iTAG demonstrates both extremely high annotation accuracy and naturalness across extensive tests, and the results of testing text-based causal discovery algorithms with the generated data show high statistical correlation with real-world data. This suggests that iTAG-generated data can serve as a practical surrogate for scalable benchmarking of text-based causal discovery algorithms.
Reinforcement learning with verifiable rewards (RLVR) has significantly boosted the reasoning capability of language models (LMs). However, existing RLVR approaches train LMs based on their own on-policy responses and are constrained by the initial capability of LMs, thus prone to exploration stagnation, in which LMs fail to solve more training problems and cannot further learn from the training data. Some approaches try to address this by leveraging off-policy solutions to training problems, but rely on external expert guidance that is limited in availability and scalability. In this work, we propose LTE (Learning to reason from Trial and Error), an approach that hints LMs with their previously self-made mistakes, not requiring any external expert guidance. Experiments validate the effectiveness of LTE, which outperforms the normal group relative policy optimization (GRPO) by 5.02 in Pass@1 and 9.96 in Pass@k on average across six mathematical reasoning benchmarks for Qwen3-8B-Base and even performs better than methods that require external guidance. Further analysis confirms that LTE successfully mitigates exploration stagnation and enhances both exploitation and exploration during training. Our code is available at https://github.com/JamyDon/LTE.
Quotation recommendation enriches writing by suggesting quotations that fit a given context, but prior systems largely focus on topical relevance and overlook what makes quotes memorable. Based on a user study, we find that preferred quotations are often unexpected yet rational, motivating the goal of selecting quotes that are contextually novel while semantically coherent. We propose NovelQR, which (1) uses a generative label agent to map quotations and contexts into multi-dimensional deep-meaning labels for label-enhanced retrieval, and (2) reranks candidates with a token-level novelty estimator that mitigates auto-regressive continuation bias. Experiments on bilingual datasets across diverse domains show that NovelQR is preferred by human judges and improves overall recommendation quality over strong baselines, while achieving competitive novelty estimation.
From Completion to Editing: Unlocking Context-Aware Code Infilling via Search-and-Replace Instruction Tuning
PDF ↗The dominant Fill-in-the-Middle (FIM) paradigm for code completion is constrained by its rigid inability to correct contextual errors and reliance on unaligned, insecure Base models. While Chat LLMs offer safety and Agentic workflows provide flexibility, they suffer from performance degradation and prohibitive latency, respectively. To resolve this dilemma, we propose Search-and-Replace Infilling (SRI), a framework that internalizes the agentic verification-and-editing mechanism into a unified, single-pass inference process. By structurally grounding edits via an explicit search phase, SRI harmonizes completion tasks with the instruction-following priors of Chat LLMs, extending the paradigm from static infilling to dynamic context-aware editing. We synthesize a high-quality dataset, SRI-200K, and fine-tune the SRI-Coder series. Extensive evaluations demonstrate that with minimal data (20k samples), SRI-Coder enables Chat models to surpass the completion performance of their Base counterparts. Crucially, unlike FIM-style tuning, SRI preserves general coding competencies and maintains inference latency comparable to standard FIM. We release our dataset and models, establishing SRI as a robust, secure, and efficient alignment recipe for next-generation interactive development.
ChessArena: A Chess Testbed for Evaluating Strategic Reasoning Capabilities of Large Language Models
PDF ↗Recent large language models (LLMs) have shown strong reasoning capabilities. However, a critical question remains: do these models possess genuine strategic reasoning, or do they primarily excel at pattern recognition? To address this, we present ChessArena, a chess-based testbed for evaluating LLMs. Chess demands strategic reasoning, precise rule adherence, and the ability to track complex game states. ChessArena is a competitive framework where LLMs play against each other under four play modes. We evaluate 13 LLMs across over 800 games, testing basic understanding, move selection, and puzzle solving. Results reveal significant shortcomings: no model beats Maia-1100 (human amateur level), and some lose to random play. We also present a strong baseline: our fine-tuned Qwen3-8B substantially improves performance, approaching much larger state-of-the-art reasoning models.