论文检索

Suyoung Bae, CheolWon Na, Jaehoon Lee, Yumin Lee, YunSeok Choi, Jee-Hyong Lee

As Large Language Models (LLMs) have become capable of generating long and descriptive code summaries, accurate and reliable evaluation of factual consistency has become a critical challenge. However, previous evaluation methods are primarily designed for short summaries of isolated code snippets. Consequently, they struggle to provide fine-grained evaluation of multi-sentence functionalities and fail to accurately assess dependency context commonly found in real-world code summaries.To address this, we propose ReFEree, a reference-free and fine-grained method for evaluating factual consistency in real-world code summaries. We define factual inconsistency criteria specific to code summaries and evaluate them at the segment level using these criteria along with dependency information. These segment-level results are then aggregated into a fine-grained score. We construct a code summarization benchmark with human-annotated factual consistency labels. The evaluation results demonstrate that ReFEree achieves the highest correlation with human judgment among 13 baselines, improving 15-18% over the previous state-of-the-art. Our code and data are available at https://github.com/bsy99615/ReFEree.git.

Yi Jing, Weiyun Qiu, Yihang Peng, Zhifang Sui

Language change both reflects and shapes social processes, and the semantic evolution of foundational concepts provides a measurable trace of historical and social transformation. Despite recent advances in diachronic semantics and discourse analysis, existing computational approaches often (i) concentrate on a single concept or a single corpus, making findings difficult to compare across heterogeneous sources, and (ii) remain confined to surface lexical evidence, offering insufficient computational and interpretive granularity when concepts are expressed implicitly. We propose HistLens, a unified, SAE-based framework for multi-concept, multi-corpus conceptual-history analysis. The framework decomposes concept representations into interpretable features and tracks their activation dynamics over time and across sources, yielding comparable conceptual trajectories within a shared coordinate system. Experiments on long-span press corpora show that HistLens supports cross-concept, cross-corpus computation of patterns of idea evolution and enables implicit concept computation. By bridging conceptual modeling with interpretive needs, HistLens broadens the analytical perspectives and methodological repertoire available to social science and the humanities for diachronic text analysis.

Clara Lachenmaier, Hannah Bultmann, Sina Zarrieß

Repair, an important resource for resolving trouble in human–human conversation, remains underexplored in human–LLM interaction.In this study, we investigate how LLMs engage in the interactive process of repair in multi-turn dialogues around solvable and unsolvable math questions. We examine whether models initiate repair themselves and how they respond to user-initiated repair. Our results show strong differences across models: reactions range from being almost completely resistant to (appropriate) repair attempts to being highly susceptible and easily manipulated. We further demonstrate that once conversations extend beyond a single turn, model behavior becomes more distinctive and less predictable across systems. Overall, our findings indicate that each tested LLM exhibits its own characteristic form of unreliability in the context of repair.

Hyeong Kyu Choi, Jerry Zhu, Sharon Li

Multi-agent debate (MAD) aims to improve large language model (LLM) reasoning by letting multiple agents exchange answers and then aggregate their opinions. Yet recent studies reveal that agents are not neutral: they are prone to identity-driven sycophancy and self-bias, uncritically adopting a peer’s view or stubbornly adhering to their own prior output, undermining the reliability of debate. In this work, we present the first principled framework that joins sycophancy and self-bias to mitigate and quantify identity bias in MAD. First, we formalize the debate dynamics as an identity-weighted Bayesian update process. Second, we propose response anonymization: by removing identity markers from prompts, agents cannot distinguish "self" from "peer", which forces equal weights on agent identity, thereby reducing bias and improving trustworthiness. Third, we define the Identity Bias Coefficient (IBC), a principled bias metric that measures an agent’s tendency to follow its peer versus itself. Empirical studies across multiple models and benchmarks confirm that identity bias is widespread, with sycophancy far more common than self-bias. Our findings highlight the need to ensure that MAD systems reason based on content rather than identity.

Kun Li, Zenan Xu, Junan Li, Zengrui Jin, Jinghao Deng, Zexuan Qiu, Bo Zhou

Tool-Integrated Reasoning has emerged as a key paradigm to augment Large Language Models (LLMs) with computational capabilities, yet integrating tool-use into long Chain-of-Thought (long CoT) remains underexplored, largely due to the scarcity of training data and the challenge of integrating tool-use without compromising the model’s intrinsic long-chain reasoning. In this paper, we introduce DART (Discovery And Reinforcement of Tool-Integrated Reasoning Chains via Rollout Trees), a reinforcement learning framework that enables spontaneous tool-use during long CoT reasoning without additional human annotation. DART operates by constructing dynamic rollout trees during training to discover valid tool-use opportunities, branching out at promising positions to explore tool-integrated trajectories. Subsequently, a tree-based process advantage estimation identifies and credits specific sub-trajectories where tool invocation positively contributes to the solution, effectively reinforcing these beneficial behaviors during training. Extensive experiments on challenging benchmarks like AIME and GPQA-Diamond demonstrate that DART significantly outperforms existing methods, successfully harmonizing tool execution with long CoT reasoning.

Bruce Qin, Dan Goldwasser

Large language models (LLMs) can both evaluate and explain text quality; however, most existing evaluators operate as static classifiers and lack the ability to refine their reasoning through interaction. We propose an Iterative Alpha–Beta Learning framework that jointly trains two complementary 8B models: an Alpha (\alpha) classifier that assesses pairwise story engagement, and a Beta (\beta) generator that produces structured, rubric-guided comparative explanations. The two models co-evolve within a closed feedback loop: \alpha provides probabilistic preference signals to guide \beta’s Direct Preference Optimization (DPO), while \beta’s improved explanations are reintegrated to retrain \alpha via a KL-based contrastive objective. This dual optimization enables mutual learning: \alpha gains interpretability and robustness from \beta’s textual rationales, while \beta acquires stronger alignment and discriminative precision from \alpha’s confidence deltas. Experiments on human-annotated story-pair datasets HANNA show that the proposed system consistently outperforms strong single-model baselines in both accuracy and explanation quality across multiple iterative rounds.

Qian Cao, Yahui Liu, Wei Bi, Yi Zhao, Ruihua Song, Xiting Wang, Ruiming Tang, Guorui Zhou, Han Li

Reinforcement learning (RL)-based enhancement of large language models (LLMs) often leads to reduced output diversity, undermining their utility in open-ended tasks like creative writing. Current methods lack explicit mechanisms for guiding diverse exploration and instead prioritize optimization efficiency and performance over diversity. This paper proposes an RL framework structured around a semi-structured long Chain-of-Thought (CoT), in which the generation process is decomposed into explicitly planned intermediate steps. We introduce a Diverse Planning Branching method that strategically introduces divergence at the planning phase based on diversity variation, alongside a group-aware diversity reward to encourage distinct trajectories. Experimental results on creative writing benchmarks demonstrate that our approach significantly improves output diversity without compromising generation quality, consistently outperforming existing baselines.

Weixi Tong, Yifeng Di, Tianyi Zhang

Existing web agents typically initiate exploration from the root URL, which is inefficient for complex websites with deep hierarchical structures. Without a global view of the website’s structure, agents frequently fall into navigation traps, explore irrelevant branches, or fail to reach target information within a limited budget. We propose Mango, a multi-agent web navigation method that leverages the website structure to dynamically determine optimal starting points. We formulate URL selection as a multi-armed bandit problem and employ Thompson Sampling to adaptively allocate the navigation budget across candidate URLs. Furthermore, we introduce an episodic memory component to store navigation history, enabling the agent to learn from previous attempts. Experiments on WebVoyager demonstrate that Mango achieves a success rate of 63.6% when using GPT-5-mini, outperforming the best baseline by 7.3%. Furthermore, on WebWalkerQA, Mango attains a 52.5% success rate, surpassing the best baseline by 26.8%. We also demonstrate the generalizability of Mango using both open-source and closed-source models as backbones. Our data and code are open-source and available at https://github.com/VichyTong/Mango.

Gaifan Zhang, Danushka Bollegala

We propose a method to compare and visualise sentence encoders at scale by creating a map of encoders where each sentence encoder is represented in relation to the other sentence encoders. Specifically, we first represent a sentence encoder using an embedding matrix of a sentence set, where each row corresponds to the embedding of a sentence. Next, we compute the PIP matrix for a sentence encoder using its embedding matrix. Finally, we create a feature vector for each sentence encoder that reflects its QRE with respect to a unit base encoder. We construct a map of encoders covering 1101 publicly available sentence encoders, providing a new perspective of the landscape of the pre-trained sentence encoders. Our map accurately reflects various relationships between encoders, where encoders with similar attributes are proximally located on the map. Moreover, our encoder feature vectors can be used to accurately infer downstream task performance of the encoders, such as in retrieval and clustering tasks, demonstrating the correctness of our map.

Lawrence Lim, Jiaming Liu, Vikas Kalagi, Divyakant Agrawal, Amr El Abbadi

A promising direction for enabling private queries to large language models (LLMs) is with homomorphic encryption (HE). An open problem is performing token sampling under HE. In this paper, we introduce Hyperion, an efficient HE algorithm for inverse transform sampling, enabling private token sampling with 1 comparison depth, O(1) amortized comparisons, and O(\log n) rotations. We implement our approach and demonstrate that it samples tokens in 0.14 seconds for 32k tokens (\approx 4.4\ \mu \mathrm{s} per token) on GPU, achieving a 100\times latency improvement over prior work.

Ziyi Yin, Yuanpu Cao, Ting Wang, Jinghui Chen, Fenglong Ma

The explainable medical coding task aims to automatically assign International Classification of Diseases (ICD) codes to clinical notes while providing explicit justifications for each assignment. Recent approaches employ large language models (LLMs) to generate such explanations. However, their performance remains limited due to a lack of understanding of the clinical meanings of ICD codes. Additionally, the vast ICD code space further complicates the task of accurate prediction. To address these challenges, we propose the ICDAGENT framework, which consists of two collaborative LLM agents: a coding agent and a critical agent. The coding agent extracts ICD codes and generates preliminary rationales, while the critical agent performs fine-grained chain-of-thought reasoning to verify and refine them. Furthermore, the critical agent is trained with a rationale-aware reward, combined with reinforcement learning, enabling it to distinguish between correct and incorrect reasoning and ensure explanation accuracy. Experiments across multiple ICD coding standards and datasets demonstrate that ICDAGENT achieves effective ICD coding with accurate and trustworthy explanations.

Francesco Ortu, Zhijing Jin, Diego Doimo, Alberto Cazzaniga

Vision-language models (VLMs) increasingly combine visual and textual information to perform complex tasks. However, conflicts between their internal knowledge and external visual input can lead to hallucinations and unreliable predictions. In this work, we investigate the mechanisms that VLMs use to resolve cross-modal conflicts by introducing WHOOPS-AHA!, a dataset of multimodal counterfactual queries that deliberately contradict internal commonsense knowledge. Through logit inspection, we identify a small set of attention heads that mediate this conflict. By intervening in these heads, we can steer the model towards its internal parametric knowledge or the visual information. Our results show that attention patterns on these heads effectively locate image regions that influence visual overrides, providing a more precise attribution compared to gradient-based methods.

Yuanfei Wang, Xinju Huang, Fangwei Zhong, Yaodong Yang, Yizhou Wang, Yuanpei Chen, Hao Dong

Effective real-world human–agent interactions, such as household robotic services, are often long-term and repeated. Beyond executing tasks, agents are expected to quickly become familiar with individual users. In everyday use, people do not want to repeatedly specify precise instructions. Instead, they prefer agents that adapt to their habits and preferences over interaction while minimizing communication effort. This poses a key challenge: enabling agents to rapidly align with user needs and provide proactive assistance within limited communication. To study this problem in a realistic embodied setting, we first introduce HA-Desire, a home assistance simulation environment. HA-Desire features an LLM-driven proxy user with value-driven preferences and natural language behavior, enabling systematic evaluation of how agents adapt to users across interactions and satisfy their desires. We further propose FAMER, a framework that integrates goal-relevant memory, desire-centered mental reasoning, and efficient communication to infer user preferences from interaction while reducing unnecessary dialogue. Experiments across embodied household tasks and different LLMs show that FAMER improves both task success and interaction efficiency compared to existing baselines, highlighting the importance of communication-efficient desire alignment for proactive embodied agents that support users without requiring frequent instructions.

David Ilić, David Stanojević, Kostadin Cvejoski

Fine-tuned language models pose significant privacy risks, as they may memorize and expose sensitive information from their training data. Membership inference attacks (MIAs) provide a principled framework for auditing these risks, yet existing methods achieve limited detection rates, particularly at the low false-positive thresholds required for practical privacy auditing. We present EZ-MIA, a membership inference attack that exploits a key observation: memorization manifests most strongly at error positions, specifically tokens where the model predicts incorrectly yet still shows elevated probability for training examples. We introduce the Error Zone (EZ) score, which measures the directional imbalance of probability shifts at error positions relative to a pretrained reference model. This principled statistic requires only two forward passes per query and no model training of any kind. On WikiText with GPT-2, EZ-MIA achieves 3.8\times higher detection than the previous state-of-the-art under identical conditions (66.3% versus 17.5% true positive rate at 1% false positive rate), with near-perfect discrimination (AUC 0.98). At the stringent 0.1% FPR threshold critical for real-world auditing, we achieve 8\times higher detection than prior work (14.0% versus 1.8%), requiring no reference model training. These gains extend to larger architectures: on AG News with Llama-2-7B, we achieve 3\times higher detection (46.7% versus 15.8% TPR at 1% FPR). These results establish that privacy risks of fine-tuned language models are substantially greater than previously understood, with implications for both privacy auditing and deployment decisions. Code is available at https://github.com/JetBrains-Research/ez-mia.

Yuxia Wang, Rui Xing, Jonibek Mansurov, Giovanni Puccetti, Zhuohan Xie, Minh Ngoc Ta, Jiahui Geng, Jinyan Su, Mervat Abassy, Saadeldine Eletter 等

Prior studies have shown that distinguishing text generated by Large Language Models (LLMs) from human-written one is highly challenging for humans, and often no better than random guessing. To verify the generalizability of this finding across languages and domains, we perform an extensive case study to identify the upper bound of human detection accuracy. Across 16 datasets covering 9 languages and 9 domains, 19 annotators achieved an average detection accuracy of 87.6%, thus challenging previous conclusions. We find that major gaps between human and machine text lie in concreteness, cultural nuances, and diversity. Prompting by explicitly explaining the distinctions in the prompts can partially bridge the gaps in over 50% of the cases. However, we also find that humans do not always prefer human-written text, particularly when they cannot clearly identify its source. We release our dataset, the human labels, and the annotator metadata at https://github.com/xnlp-lab/HumanEval-MGT.

Ruochong Xiong, Qien Li, Wangwang Lian, Yulong Wan, Hanlin Xue, Zhouxing Tan, Han Yang, Fengyu Lu, Junfei Liu

The widespread deployment of large language models (LLMs) makes detecting LLM-Generated text a critical security task. Existing methods, primarily relying on output probabilities from proxy models or single semantic features, suffer from distribution misalignment and limited interpretability. We observe that machine-generated text exhibits a directionally consistent systematic translation relative to human-written text within the joint semantic-structural space. Accordingly, we propose ProSSD, a statistical framework utilizing supervised subspace learning to extract compact features and construct conditional semantic distributions based on syntactic structures. By employing a likelihood ratio test, we derive a modified Mahalanobis distance, weighted by the Wasserstein distance, as the discriminative metric. Experiments demonstrate ProSSD’s superior robustness and computational efficiency across cross-domain, cross-model, and adversarial scenarios. Furthermore, we reveal the phenomena of systematic semantic translation and semantic collapse in machine-generated text, offering interpretable statistical insights into LLM generation behaviors.

Chen Zhan, Xiaoyu Tan, Gengchen Ma, Yu-Jie Xiong, Xiaoyan Jiang, Xihe Qiu

The integration of Large Language Models (LLMs) into clinical decision support is critically obstructed by their opaque and often unreliable reasoning. In the high-stakes domain of healthcare, correct answers alone are insufficient; clinical practice demands full transparency to ensure patient safety and enable professional accountability. A pervasive and dangerous weakness of current LLMs is their tendency to produce "correct answers through flawed reasoning." This issue is far more than a minor academic flaw; such process errors signal a fundamental lack of robust understanding, making the model prone to broader hallucinations and unpredictable failures when faced with real-world clinical complexity. In this paper, we establish a framework for trustworthy clinical argumentation by adapting the Toulmin model to the diagnostic process. We propose a novel training pipeline: Curriculum Goal-Conditioned Learning (CGCL), designed to progressively train LLM to generate diagnostic arguments that explicitly follow this Toulmin structure. CGCL’s progressive three-stage curriculum systematically builds a solid clinical argument: (1) extracting facts and generating differential diagnoses; (2) justifying a core hypothesis while rebutting alternatives; and (3) synthesizing the analysis into a final, qualified conclusion. We validate CGCL using T-Eval, a quantitative framework measuring the integrity of the diagnosis reasoning. Experiments show that our method achieves diagnostic accuracy and reasoning quality comparable to resource-intensive Reinforcement Learning (RL) methods, while offering a more stable and efficient training pipeline.

Chi-Min Chan, Yujin Zhou, Pengcheng Wen, Boqin Yin, Jiaming Ji, Juntao Dai, Wei Xue, Sirui Han, Yike Guo

The rise of Omni-modality Large Language Models (OLLMs) capable of jointly processing text, audio, and visual inputs marks a major step toward general intelligence. Ensuring their alignment with human preferences requires effective Omni-modality Reward Models (ORMs), which serve as surrogates for human judgment to guide OLLMs behavior. However, ORMs evaluation remains underdeveloped in the previous literature. Existing benchmarks are largely text-centric or limited to bimodal tasks, restricting comprehensive assessment for ORMs. To bridge this gap, we introduce Omni-RewardBench, the first benchmark for comprehensive evaluation of ORMs across modalities. In short, our contributions are threefold: (1) a hybrid automatic-annotation and human-verification pipeline to construct high-quality evaluation data; (2) extensive experiments on 20+ models, including inherently omni-modal and modality-bridged systems. Our experimental results demonstrate that current OLLMs fall short as reward models, revealing several common failure modes such as perception failure, modality dominance failure, and cross-modal fusion failure; and (3) strong correlations between Omni-RewardBench scores and downstream performance (IID r = 0.94, OOD r = 0.72), validating its reliability as a predictor of real-world capability and alignment quality.

Katelyn X. Mei, Yi-Li Hsu, Minjoon Choi, Zongwan Cao, Chenjun Xu, Bingbing Wen, Su Lin Blodgett, Lucy Lu Wang

Human evaluation plays a critical role in assessing the quality of generated text. However, the reliability and reproducibility of these evaluations depend on transparent and well-documented protocols—details that are frequently missing in current practice. In this work, we conduct a large-scale analysis of human evaluation protocols for evaluating long-form generation tasks in *CL conference publications from 2023–2025, including a full manual review of 284 papers and LLM-assisted analysis for another 1.8k+ papers. We define a set of 20 reportable criteria related to reproducibility of human evaluation studies, and apply these criteria to systematically examine reporting norms and practices within the community. We find widespread under-reporting of important aspects of human evaluation study design, leading to ambiguity about what was measured and how, who contributed judgments, and how judgments should be interpreted. Based on these findings, we outline actionable recommendations to support more transparent and reproducible reporting in future research.

Disen Liao, Freda Shi

Tokenization is the first step in every language model (LM), yet it never takes the sounds of words into account. We investigate how tokenization influences text-only LMs’ ability to represent phonological knowledge. Through a series of probing experiments, we show that subword-based tokenization systematically weakens the encoding of both local (e.g., rhyme) and global (e.g., syllabification) phonological features. To quantify this effect, we introduce the syllabification-tokenization alignment distance (STAD), a metric that measures the misalignment between a model’s tokenization and the natural syllable boundaries of words, and find that higher misalignment correlates with poorer phonological representations, providing a simple diagnostic for phonology-aware tokenization. To address these limitations, we propose a lightweight IPA-based fine-tuning method that infuses phonological awareness into LMs, leading to consistent improvements across three phonology-related tasks while largely preserving math and general reasoning ability, with 1.1% and 0.9% drops on GSM8K and MMLU, respectively.[Our code is available at <https://github.com/liaodisen/Tokenization-Phonology>]