论文检索

Ivan Rodkin, Daniil Orel, Konstantin Smirnov, Arman Bolatov, Bilal Elbouardi, Besher Hassan, Yuri Kuratov, Aydar Bulatov, Preslav Nakov, Timothy Baldwin 等

Reasoning is a core capability of large language models (LLMs), yet how multi-step reasoning is learned and executed remains unclear. We study this question in a controlled cellular-automata (1dCA) framework that excludes memorization by using disjoint training and test rules. Given a short state sequence, the model is required to infer the hidden local rule and then chain it to predict multiple future steps. Our evaluation shows that LLMs largely fail to reliably solve a natural-language proxy of the proposed task. We find that most neural architectures trained from scratch can learn rule inference and achieve high next-step accuracy, but performance drops sharply as the required number of intermediate reasoning steps increases. Experiments show that increasing model depth is crucial, and extending effective depth via recurrence, memory, or test-time compute improves results but remains bounded. Code is available on github: https://github.com/RodkinIvan/associative-recurrent-memory-transformer/tree/ACT.

Seungone Kim, Ian Wu, Jinu Lee, Xiang Yue, Seongyun Lee, Minkyeong Moon, Carolin Lawrence, Kiril Gashteovski, Julia Hockenmaier, Graham Neubig 等

Language model (LM) evaluators that generate chain-of-thought (CoT) reasoning are widely used for the assessment of LM responses. Simultaneously, increasing LMs’ "thinking" time through scaling test-time compute has proven to be an effective technique for solving challenging problems in domains such as math and code. This raises a natural question: can an LM’s evaluation capability also be improved by scaling test-time compute? To answer this, we investigate employing reasoning models - LMs that natively generate long CoT reasoning - as evaluators. We explore scaling evaluation-time compute by using reasoning models to evaluate both the overall candidate response (i.e., outcome evaluation) and the individual reasoning steps within it (i.e., process evaluation). We observe that evaluator performance improves monotonically with the number of reasoning tokens generated, mirroring trends seen in LM reasoning. Furthermore, we use these more accurate evaluators to rerank multiple generations, and demonstrate that spending more compute at evaluation time can be as effective as increasing compute during generation for improving an LM’s problem-solving performance.

Sonal Kumar, Sreyan Ghosh, Yueqian Lin, S Sakshi, Ashish Seth, Yiran Chen, Ramani Duraiswami, Dinesh Manocha

Large Audio Language Models have shown impressive performance on single-clip audio language tasks such as automatic speech recognition, captioning, and sound event recognition. Yet, their ability to reason over interleaved multi-audio contexts-where answering a query requires relating information across multiple audio clips-remains limited. We present PolyAudio, a LALM built on Audio Flamingo 3 that targets multi-audio understanding via instruction tuning rather than massive-scale pre-training, and PolyAudio-Instruct, a high-quality instruction-tuning dataset consisting of 1.3M+ QA pairs, spanning over 14 task subsets to empower multi-audio understanding and reasoning. PolyAudio uses an explicit interleaved representation with clip indexing to encourage faithful grounding and reduce ambiguity in multi-clip references. We evaluate PolyAudio on a diverse suite of multi-audio benchmarks alongside standard single-audio tasks. PolyAudio achieves strong performance on multi-audio reasoning, outperforming competitive baselines that are also often limited to reasoning over up-to 2 audio clips, while preserving robust single-clip performance. Overall, our results suggest that precise, academic-scale multi-audio instruction tuning can unlock advanced cross-clip reasoning capabilities, enabling more capable audio-centric assistants.

Ying Liu, Dong Shuai, Cui Zibo, TengQi Ye, Gang Wu

Relation extraction is a fundamental task in information extraction. Still, existing supervised approaches rely heavily on large-scale annotated data, limiting their applicability in domain-specific and low-resource scenarios. Prompt-based methods with large language models provide a parameter-efficient alternative; however, their performance is susceptible to prompt design, which often requires extensive domain expertise and heuristic trial-and-error. We propose REPO, a reinforcement learning-based automated prompt optimization framework for domain relation extraction. REPO formulates prompt construction as a structured, sequential decision-making problem, optimizing prompt quality through interaction with a black-box LLM. To enable efficient and stable optimization, we introduce a two-stage framework comprising an initial prompt-construction stage that generates semantically grounded candidates and a DRL-based refinement stage that iteratively improves prompts within a constrained, domain-aware action space. We further design a composite evaluation metric that integrates extraction accuracy and semantic consistency to serve as a dense reward signal. Extensive experiments on multiple relation extraction datasets across medical, financial, legal, and news domains demonstrate that REPO consistently outperforms existing prompt-based methods and supervised baselines. Ablation studies further confirm the effectiveness and robustness of the proposed DRL-based prompt optimization strategy. Our code is available at https://github.com/dddong2-star/REPO.

Yu Zhang, Peter Belcak, Shizhe Diao, Yonggan Fu, Shaona Ghosh, Morteza Mardani, Eileen Margaret Peters Long, Bei Yu, Pavlo Molchanov

Hallucination detection remains a significant challenge for large language models. Existing agentic applications rely on LLMs to self-assess the factuality of their outputs using single-step “LLM-as-a-judge” prompts. However, even when equipped with ground truth information, current LLMs still fall short in detecting hallucinations, and this one-shot evaluation offers neither the transparency nor the granularity needed to diagnose where and why the detection fails. To address this gap, we introduce PROBE (Process-based Benchmark for Hallucination Detection), a comprehensive benchmark that breaks down hallucination detection into four critical steps: claim decomposition, evidence finding, evidence evaluation, and hallucination localization, and evaluates each step individually. PROBE consists of 12,000 test cases across three task types—summarization, question answering, and style transfer. Critically, we demonstrate that when hallucination detection is treated as a multi-step process, all models achieve considerably better performance. Through extensive evaluation, we show that current LLMs struggle chiefly with evidence finding, and that finetuning on our released training data substantially improves performance on this step. PROBE represents a significant step toward more transparent, diagnosable, and robust hallucination detection systems.

Zichuan Fu, Xian Wu, Guojing Li, Yejing Wang, Yijun Chen, Zhao Zihao, Luo Yixuan, Hanyu Yan, Yefeng Zheng, Xiangyu Zhao

Recent advancements in large language models (LLMs) have catalyzed the rise of reasoningintensive inference paradigms, where models perform explicit step-by-step reasoning before generating final answers. While such approaches improve answer quality and interpretability, they incur substantial computational overhead due to the prolonged generation sequences. In this paper, we propose Tandem, a novel collaborative framework that synergizes large and small language models (LLMs and SLMs) to achieve high-quality reasoning with significantly reduced computational cost. Specifically, the LLM serves as a strategic coordinator, efficiently generating a compact set of critical reasoning insights. These insights are then used to guide a smaller, more efficient SLM in executing the full reasoning process and delivering the final response. To balance efficiency and reliability, Tandem introduces a cost-aware termination mechanism that adaptively determines when sufficient reasoning guidance has been accumulated, enabling early stopping of the LLM’s generation. Experiments on mathematical reasoning and code generation benchmarks demonstrate that Tandem reduces computational costs by approximately 40% compared to standalone LLM reasoning, while achieving superior or competitive performance. Furthermore, the sufficiency classifier trained on one domain transfers effectively to others without retraining. The code is available at: https://github.com/Applied-MachineLearning-Lab/ACL2026_Tandem.

Jiayuan Guo, Yueyang Su, Yuyao Ge, Saiping Guan, Lei Yu, Jiafeng Guo, Xueqi Cheng

Long context large language models exhibit the “lost in the middle” problem, where models struggle to effectively utilize information located in the middle of long contexts. Although existing workflow-based long context methods (e.g., RAG) alleviate this problem and perform well on specific datasets, can their effectiveness generalize to all types of datasets? In this work, we systematically investigate the cross-dataset generalization of long context methods. Our evaluation reveals that these methods are not universally effective. Such substantial performance variability underscores the risks of performance degradation associated with the indiscriminate application of long context methods. We investigated the reason for the failure of long context methods. We found that the intrinsic decomposition mechanisms of long context methods hinder context dependency modeling, causing these methods to suffer performance declines on documents with strong context dependency. To address this issue, We propose CoDaR (**Co**ntext **D**ependency-**a**ware **R**outing), a training-free adaptive routing strategy. By analyzing the context dependency strength of documents, CoDaR adaptively invokes long context methods, thereby significantly enhancing their overall robustness across different types of datasets.

Eunike Andriani Kardinata, Yusuke Sakai, Taro Watanabe

Acceptability judgments provide a crucial basis for understanding how sentences are perceived as natural or well-formed, and they are increasingly used to assess the linguistic capability of large language models (LLMs). Unlike grammaticality, acceptability depends not only on structural form but also on contextual and domain-specific factors. Most prior work evaluates sentences in isolation, and relatively little is known about how explicit contextual cues influence LLM acceptability judgments across domains. This study examines how contextual information affects model-generated acceptability ratings across multiple domains and several LLMs, using different forms of domain-specific contextual cues to situate sentences in their intended usage settings. The results show that context can meaningfully shift model judgments, although its effects vary across models and domains. Overall, the findings provide evidence on contextual effects in LLM acceptability judgment and support the development of more context-aware evaluation frameworks.

Hao Pang, Changcheng Li, Yingxue Liu

Language evolution is cognitively motivated by the reduction of communicative effort. Current research exploring this reported tendency has been constrained by the heavy reliance on manually annotated resources (e.g., dependency parsing) as well as a narrow focus (e.g., syntax as the single metric). To transcend these limitations, we propose two measures: Attention-based Structural Distance (ASD) and Semantic Space Distance (SSD). ASD is a parser-free measure of syntactic locality derived from the attention mechanism of pretrained large language models (LLM), while SSD is a measure of lexical distances that quantify the degree of separation between different parts of speech in the word vector space. Based on multiple diachronic and multilingual corpora, our experiments show a significant decrease of ASD while an increase of SSD, which implies a language developmental trend towards structural compactness and semantic divergence. Our research pioneers a novel lens grounded in LLM for studying language evolution, which has two major contributions. Linguistically, our study corroborates the hypothesized law of human language evolution by demonstrating that its development optimizes syntactic locality as well as functional semantic discriminability. Cognitively, our study shows that human and LLMs share common characteristics in language processing, lending support to the potential of employing LLMs in the study of human cognition.

Ijun Jang, Jewon Yeom, Juan Yeo, Hyunggyu Lim, Taesup Kim

Knowledge distillation (KD) is a widely adopted technique for transferring capabilities from large language models to smaller student models. However, conventional supervised KD often suffers from a distribution mismatch between training and inference. While on-policy KD approaches attempt to mitigate this issue by learning directly from student-generated outputs, they frequently encounter training instabilities and noisy teacher feedback during early optimization stages. These challenges manifest as pathological gradients in forward KL objectives when students encounter unfamiliar tokens, or as a collapse in distributional diversity within reverse KL regimes. To address these limitations, we propose Veto, an objective-level reformulation that constructs a geometric target distribution in logit space to emphasize agreement between the teacher and the student. By introducing a tunable parameter \beta, Veto serves as an Adaptive Gradient Veto that stabilizes optimization by suppressing harmful gradients on low-confidence tokens, while simultaneously acting as a Decisiveness Knob to balance reward-driven performance with output diversity. Extensive experiments across various reasoning and generation tasks demonstrate that Veto consistently outperforms supervised fine-tuning and existing on-policy baselines.

Aakriti Agrawal, Mucong Ding, Chenghao Deng, Zora Che, Arjun Rajaram, Anirudh Satheesh, Bang An, C. Bayan Bruss, John Langford, Furong Huang

With Large Language Models (LLMs) rapidly approaching and potentially surpassing human-level performance, it has become imperative to develop approaches capable of effectively supervising and enhancing these powerful models using smaller, human-level models exposed to only human-level data. We address this critical weak-to-strong (W2S) generalization challenge by proposing a novel method aimed at improving weak experts, by training on the same limited human-level data, enabling them to generalize to complex, super-human-level tasks. Our approach, called EnsemW2S, employs a token-level ensemble strategy that iteratively combines multiple weak experts, systematically addressing the shortcomings identified in preceding iterations. By continuously refining these weak models, we significantly enhance their collective ability to supervise stronger student models. We extensively evaluate the generalization performance of both the ensemble of weak experts and the subsequent strong student model across in-distribution (ID) and out-of-distribution (OOD) datasets. For OOD, we specifically introduce question difficulty as an additional dimension for defining distributional shifts. Our empirical results demonstrate notable improvements, achieving 4%, and 3.2% improvements on ID datasets and, upto 6% and 2.28% on OOD datasets for experts and student models respectively, underscoring the effectiveness of our proposed method in advancing W2S generalization.

Sagnik Anupam, Alexander Shypula, Osbert Bastani

Recent work has demonstrated the potential of large language models (LLMs) for program optimization, a key challenge in programming languages. We propose a blackbox adaptation method called Retrieval Augmented Search (RAS) that performs beam search over candidate optimizations; at each step, it retrieves in-context examples from a given training dataset of slow-fast program pairs to guide the LLM. Critically, we find that performing contextual retrieval based on an LLM-generated natural language description significantly outperforms retrieval based on the source code. We also propose AEGIS, a method for improving interpretability by decomposing training examples into "atomic edits” that are significantly more incremental in nature. We show that RAS performs up to 2.06\times better than prior state-of-the-art blackbox adaptation strategies on optimizing C++ programs, and that AEGIS performs up to 1.37\times better while making significantly smaller edits. We also show that using RAS improves the mean runtime percentile of Python programs by 10.27 compared to baselines.

Weixiang Zhao, Yichen Zhang, Yingshuo Wang, Yang Deng, Yanyan Zhao, Xuda Zhi, Yongbo Huang, Hao He, Wanxiang Che, Bing Qin 等

Experience-driven self-evolution has emerged as a promising paradigm for improving the autonomy of large language model agents, yet its reliance on self-curated experience introduces underexplored safety risks. In this study, we investigate how experience accumulation and utilization in self-evolving agents affect safety performance across web-based and embodied environments. Notably, experience gathered solely from benign tasks can still compromise safety in high-risk scenarios. Further analysis attributes this degradation to the execution-oriented nature of accumulated experience, which reinforces agents’ tendency to act rather than refuse. In more realistic settings where agents encounter both benign and harmful tasks, refusal-related experience mitigates safety decline but induces over-refusal, revealing a fundamental safety–utility trade-off. Overall, our findings expose inherent limitations of current self-evolving agents and call for more principled strategies to ensure safe and reliable adaptation.

Wei Fang, James R. Glass

LLM agents operating over massive, dynamic tool libraries rely on effective retrieval, yet standard single-shot dense retrievers struggle with complex requests. These failures primarily stem from the disconnect between abstract user goals and technical documentation, and the limited capacity of fixed-size embeddings to model combinatorial tool compositions. To address these challenges, we propose ToolQP, a lightweight framework that models retrieval as iterative query planning. Instead of single-shot matching, ToolQP decomposes instructions into sub-tasks and dynamically generates queries to interact with the retriever, effectively bridging the semantic gap by targeting the specific sub-tasks required for composition. We train ToolQP using synthetic query trajectories followed by optimization with Reinforcement Learning with Verifiable Rewards (RLVR). Experiments demonstrate that ToolQP achieves state-of-the-art performance, exhibiting superior zero-shot generalization, robustness across diverse retrievers, and significant improvements in downstream agentic execution.

Pengyuan Qin, Linnan Tu, Yuhan Ke, Hefei Ling

Large language models (LLMs) watermarking has been proposed as an active approach for content provenance verification, yet existing evaluations are largely confined to fixed entropy settings. In this paper, we introduce EntroBench, a benchmark for LLM watermarking that systematically covers three entropy levels and seven representative tasks. We conducted a fair evaluation of eight watermarking methods through hyper-parameter search based on an anchored dataset. We find that current approaches struggle to perform consistently across different entropy levels. Our analysis reveals a clear trade-off between watermark detectability and downstream output quality that varies across tasks and entropy conditions. Furthermore, we assess watermark robustness under realistic user interaction scenarios and show that common, non-adversarial user behaviors can substantially degrade watermark signals. These results indicate that practical usage-driven perturbations pose a significant challenge to current watermarking techniques. EntroBench provides a unified evaluation framework for studying these issues and supports the development of more adaptive and robust LLM watermarking methods. Dataset and codes are available at https://github.com/py-qin/EntroBench.

Hong Kyu Lee, Ruixuan Liu, Li Xiong

Machine unlearning is an emerging technique that removes the influence of a subset of training data (forget set) from a model without full retraining, with applications including privacy protection, content moderation, and model correction. The key challenge lies in achieving strong unlearning efficacy while preserving the overall utility. Existing unlearning methods for large language models (LLMs) often rely on auxiliary models, retain datasets, or even commercial AI services. However, dependence on these external resources is often impractical and could potentially introduce additional privacy risks. In this work, we propose direct token optimization (DTO), a self-contained unlearning approach for LLMs that directly optimizes the token-level objectives to unlearn specific sequences without external resources.For each sequence to be unlearned, we identify target tokens that encode critical knowledge for unlearning and treat remaining tokens as non-target ones for maintaining the model utility. DTO maximizes an unlearning objective on target tokens and applies a utility-preservation regularizer on non-target tokens.Across multiple unlearning benchmarks, DTO improves the forget quality up to 16.8\times over the latest baselines while maintaining comparable model utility. Our code is available at github.com/Emory-AIMS/direct_token_optimization.

Pardis Sadat Zahraei, Xiaoning Wang, Nimet Beyza Bozdag, Gokhan Tur, Dilek Hakkani-Tür

Large Language Models (LLMs) are increasingly used as judges to evaluate text quality, moderate content, and assess arguments. We investigate whether alignment-instilled prior beliefs bias LLM judgments, using persuasion evaluation as a representative task. We find a systematic failure: models conflate their trained beliefs with rhetorical quality, rating identical claims differently based on belief alignment rather than argumentative merit. A bare assertion aligned with training receives higher scores than a well-crafted counter-argument, even when explicitly instructed to judge rhetoric alone. We introduce ConvinceQA, a dataset of 27,756 persuasive arguments with controlled stance variation across subjective, harmful, and misinformation domains, and demonstrate this prior prejudice across models. We exploit this failure through persuasion-based probing: evaluating minimal pairs that differ only in the subject token bypasses learned refusals and reveals hidden biases. Analysis identifies three failure modes, with belief-conditioned rating inflation accounting for 88% of cases. Cross-task validation on essay quality assessment and debate judging confirms this is a pervasive limitation.

Aakriti Agrawal, Gouthaman KV, Rohith Aralikatti, Gauri Jagatap, Jiaxin Yuan, Sarvesh Baskar, Vijay Kamarshi, Andrea Fanelli, Furong Huang

Hallucinations in Large Vision-Language Models (LVLMs) remain a persistent challenge, often stemming from inadequate integration of visual information during multimodal reasoning. A key cause is the model’s over-reliance on textual priors and underutilization of visual cues, leading to outputs that are linguistically fluent but visually inaccurate. For example, given an image of an empty kitchen countertop, an LVLM might hallucinate a “bowl of fruit” or “cup of coffee,” relying on language associations rather than visual evidence. Most LVLMs incorporate visual features by appending them to the input stream of a pre-trained LLM and training on large-scale vision-language datasets. Our systematic analysis reveals that this strategy often leads to over-dependence on textual information due to the inherent bias of LLMs towards language-dominant representations. This imbalance skews attention towards the text over visual content, weakening the model’s ability to ground outputs in visual inputs. To address this, we propose a simple yet effective visual feature incorporation method that encourages the model to learn visually-informed textual embeddings distinct from those of the base LLM and promotes a more balanced attention distribution. Experimental results across multiple hallucination benchmarks demonstrate that our method significantly reduces hallucinations and fosters more balanced multimodal reasoning. Notably, our approach achieves substantial gains, including \textbf{+9.33}% on MMVP-MLLM, \textbf{+2.99}% on POPE-AOKVQA, up to \textbf{+3.4}% on Merlin, and \textbf{+3}% on the hard-data split of HallusionBench.

Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Soeun Kim, Seungju Han, Youngjae Yu, Albert No

Machine unlearning aims to remove “forget” data while preserving knowledge from the “retain” data, yet a fundamental question arises when the two share content. By definition, an unlearned model should be indistinguishable from a model retrained solely on the retain set, which implies that shared knowledge must remain while only forget-specific content is removed. To evaluate this requirement, we introduce DUSK, the first benchmark for unlearning under realistic knowledge overlap. DUSK constructs documents containing both shared and unique knowledge and defines seven metrics to test whether methods erase forget-specific expressions without discarding shared facts. Evaluating nine recent approaches, we find that although surface text is often removed, current methods struggle to distinguish shared from unique knowledge, either erasing information that should be retained or failing to fully forget target content. DUSK provides a controlled, reproducible testbed for diagnosing these failures and guiding precise unlearning algorithms.

Qipeng Xie, Zi Liang, Jiafei Wu, Yufei Chen, Weizheng Wang, Wenao Ma, Zhong Ming, Haiqin Yang, Kaishun Wu

Large Language Models (LLMs) often exhibit extreme sensitivity to surface-level prompt variations, where minor lexical perturbations trigger disproportionate performance fluctuations. Moving beyond black-box optimization or coarse-grained templates, we conduct the first analysis of n-gram token-level mechanisms, leveraging a large-scale dataset of 132,000 prompt variants. Our investigation uncovers the Scaling Law of Prompt Performance Stability: higher average performance is inherently associated with lower variance and greater stability. We identify that this robustness is driven by two linguistic pillars: Domain-Specific Terminology, which anchors semantic boundaries, and Explicit Action Directives, which formalize reasoning trajectories. By narrowing the model’s interpretative space, these patterns effectively "lock" the generation process. We operationalize these findings into an automated Prompt-Refining Agent that autonomously restructures queries via domain anchoring and operational constraints. Empirical results show a 40.7% reduction in performance variance for code generation, offering a statistically grounded framework for robust prompt engineering.