论文检索

Zhihao Zhang, Liting Huang, Guanghao Wu, Preslav Nakov, Heng Ji, Usman Naseem

Safety alignment in Large Language Models is critical for healthcare; however, reliance on binary refusal boundaries often results in over-refusal of benign queries or unsafe compliance with harmful ones. While existing benchmarks measure these extremes, they fail to evaluate Safe Completion: the model’s ability to maximise helpfulness on dual-use or borderline queries by providing safe, high-level guidance without crossing into actionable harm. We introduce Health-ORSC-Bench, the first large-scale benchmark designed to systematically measure Over-Refusal and Safe Completion quality in healthcare. Comprising 31,920 benign boundary prompts across seven health categories (e.g., self-harm, medical misinformation), our framework uses an automated pipeline with human validation to test models at varying levels of intent ambiguity. We evaluate 30 state-of-the-art LLMs, including GPT-5 and Claude-4, revealing a significant tension: safety-optimised models frequently refuse up to 80% of "Hard" benign prompts, while domain-specific models often sacrifice safety for utility. Our findings demonstrate that model family and size significantly influence calibration: larger frontier models (e.g., GPT-5, Llama-4) exhibit "safety-pessimism" and higher over-refusal than smaller or MoE-based counterparts (e.g., Qwen-3-Next), highlighting that current LLMs struggle to balance refusal and compliance. Health-ORSC-Bench provides a rigorous standard for calibrating the next generation of medical AI assistants toward nuanced, safe, and helpful completions. Our code and data is available at: https://github.com/ZhihaoZhang97/Health-ORSC-Bench. Warning: Some contents may include toxic or undesired contents.

Thanmay Jayakumar, Deepon Halder, Raj Dabre

Cross-lingual transfer in NLP is often hindered by the "script barrier" where differences in writing systems inhibit transfer learning between languages. Transliteration, the process of converting the script, has emerged as a powerful technique to bridge this gap by increasing lexical overlap. This paper provides a comprehensive survey of the application of transliteration in cross-lingual NLP. We present a taxonomy of key motivations to utilize transliterations in language models, and also provide an overview of different approaches of incorporating transliterations as input. We analyze the evolution and effectiveness of these methods, discussing the critical trade-offs involved, and contextualize their need in modern LLMs. The review explores various contexts how transliteration is beneficial, including handling code-mixed text, leveraging language family relatedness, and pragmatic gains in inference efficiency. Based on this analysis, we provide concrete recommendations for researchers on selecting and implementing the most appropriate transliteration strategy based on their specific language, task, and resource constraints.

Kai Wang, Haoyang You, Yang Zhang, Zhongjie Wang

A core challenge for faithful LLM role-playing is sustaining consistent characterization throughout long, open-ended dialogues, as models frequently fail to recall and accurately apply their designated persona knowledge without explicit cues. To tackle this, we propose the Memory-Driven Role-Playing paradigm. Inspired by Stanislavski’s "emotional memory” acting theory, this paradigm frames persona knowledge as the LLM’s internal memory store, requiring retrieval and application based solely on dialogue context, thereby providing a rigorous test of depth and autonomous use of knowledge. Centered on this paradigm, we contribute: (1) MREval, a fine-grained evaluation framework assessing four memory-driven abilities—Anchoring, Selecting, Bounding, and Enacting; (2) MRPrompt, a prompting architecture that guides structured memory retrieval and response generation; and (3) MRBench, a bilingual (Chinese/English) benchmark for fine-grained diagnosis. The novel paradigm provides a comprehensive diagnostic for four-stage role-playing abilities across 12 LLMs. Crucially, experiments show that MRPrompt allows small models (e.g., Qwen3-8B) to match the performance of much larger closed-source LLMs (e.g., Qwen3-Max and GLM-4.7), and confirm that upstream memory gains directly enhance downstream response quality, validating the staged theoretical foundation.

Jiu Sha, Mengxiao Zhu

Aligning LLMs in low-resource multilingual settings faces a fundamental reward bottleneck: scalar rewards lack cultural generalization, while unstructured critiques remain noisy and unverifiable. To bridge this gap, we introduce a Structured Multilingual Reward Modeling Framework that extends Reinforcement Learning with Verifiable Rewards (RLVR) to subjective and open-ended tasks. The framework unifies three core components to transform abstract quality into concrete supervision: (1) a Structured Checklist Schema decomposing evaluation into granular universal reasoning steps and task-specific criteria; (2) Structured Generative Critique Modeling, which produces rubric-aligned critiques with grounded justifications; and (3) Adaptive Multilingual Reward Optimization, integrating reasoning quality and language consistency into a verifiable objective. We integrate this framework into a bootstrapped Group Relative Policy Optimization pipeline, augmented by length-aware normalization and variance stabilization to ensure stability. Extensive experiments on a newly constructed suite covering 7 subjective task categories across 50 low-resource languages demonstrate that this checklist-driven approach yields substantial improvements in reasoning capability and response quality, particularly in settings where traditional reward models exhibit significant degradation. We publicly release our models and the corresponding evaluation benchmark to facilitate further research. Our code is available at https://github.com/Shajiu/SGCM.

Zheng He, Yiwei Wang, Hongxing Wang, Yujun Cai

Large Vision-Language Models (LVLMs) confront an escalating threat from sophisticated multimodal jailbreak attacks. However, existing defense strategies suffer from three critical limitations: (1) the neglect of visual threats; (2) a lack of fine-grained specificity regarding specific attack semantics; and (3) the absence of a dedicated jailbreak detection mechanism, which leads to unnecessary defensive measures against benign inputs. To address these limitations, we propose ReCon, a novel black-box defense framework. ReCon integrates a diffusion-based image purifier to neutralize visual perturbations and an autoencoder-based detector for anomaly filtration. At its core, it employs a Reverse Safety Concept Injection module that maps detected unsafe concepts to fine-grained, constructive Safe Concepts, generating targeted prompts to precisely rectify attack semantics. Extensive experiments demonstrate that ReCon significantly enhances the robustness of LVLMs against jailbreak attacks while preserving performance on benign tasks. Disclaimer: Samples in this paper may be harmful and cause discomfort.

Yang Lyu, Jin Cao, Yang Xiao, Zhe Sun, Ben Niu, Fenghua Li, Hui LI

Distributed LLM inference avoids sending raw inputs by transmitting intermediate hidden states, a practice widely assumed to preserve privacy. We challenge this assumption and demonstrate that intermediate representations alone are sufficient to leak sensitive user attributes. This setting poses a fundamental obstacle for existing attribute inference attacks, which typically rely on auxiliary embedding-attribute pairs. To characterize this previously underexplored privacy risk, we reformulate attribute inference as zero-shot matching over candidate attributes directly in the intermediate representation space, and introduce a purely intermediate-representation-based attribute inference attack, termed IR-AIA. To address two structural challenges that hinder attribute inference from intermediate representations, we propose SG-APCR to address layer-dependent anisotropy in intermediate embeddings and a sliding-window similarity matching strategy to handle subword-level semantic fragmentation. Experiments across three LLMs and three real-world datasets show that sensitive attributes can be reliably inferred using only intermediate representations, achieving Top-1 accuracy of up to 0.997 on CMS, 0.980 on Skytrax, and 0.986 on ECHR. These results reveal that intermediate states commonly considered safe to share can expose sensitive personal attributes on their own.

Xukai Wang, Xuanbo Liu, Mingrui Chen, Haitian Zhong, Xuanlin Yang, Bohan Zeng, Jinbo Hu, Hao Liang, Junbo Niu, Xuchen Li 等

With the advancement of powerful large-scale reasoning models, effectively evaluating the reasoning capabilities of these models has become increasingly important. However, existing benchmarks designed to assess the reasoning abilities of large models tend to be limited in scope and lack the flexibility to adapt their difficulty according to the evolving reasoning capacities of the models. To address this, we propose MorphoBench, a benchmark that incorporates multidisciplinary questions to evaluate the reasoning capabilities of large models and can adjust and update question difficulty based on the reasoning abilities of advanced models. Specifically, we curate the benchmark by selecting and collecting complex reasoning questions from existing benchmarks and sources such as Olympiad-level competitions. Additionally, MorphoBench adaptively modifies the analytical challenge of questions by leveraging key statements generated during the model’s reasoning process. Furthermore, it includes questions generated using simulation software, enabling dynamic adjustment of benchmark difficulty with minimal resource consumption. We have gathered over 1,300 test questions and iteratively adjusted the difficulty of MorphoBench based on the reasoning capabilities of models such as GPT-5 and Gemini-3-Pro. MorphoBench enhances the comprehensiveness and validity of model reasoning evaluation, providing reliable guidance for improving both the reasoning abilities and scientific robustness of large models.

Zeyuan Chen, Ziqing Yang, Yihan Ma, Michael Backes, Yang Zhang

As academic submissions grow, the traditional peer review process struggles to keep up, raising concerns about quality and fairness.A trend of using large language models (LLMs) for assistance has emerged.In this work, we take a critical step toward improving the quality of LLM-generated reviews.We propose the PeerCheck framework, which investigates LLM-human review differences (\textbf{RQ1}) and explores methods to increase LLM-human similarity (\textbf{RQ2}).We first analyzed the human-written reviews with reviews generated by GPT-4o, Claude-3.7-Sonnet, and DeepSeek-V3 and found that LLMs and humans focus on different terms, e.g., LLMs prioritize theory while humans emphasize methodology and experiments.We further adopt prompt engineering, such as Chain-of-Thought (CoT), and utilize retrieval-augmented generation (RAG) to enhance the LLM-generated reviews towards human-level quality.We find CoT significantly improves the human similarity of LLM reviews, while we also discover an unexpected “RAG paradox,” i.e., experiments with RAG produce different results for various LLMs and, in some cases, even reduce review quality.Our comprehensive analysis of LLM-generated academic reviews illustrates both possibilities and limitations, contributing to a more effective, human-aligned review system.

Hao Li, Yubing Ren, Yanan Cao, Yingjie Li, Fang Fang, Shi Wang, Li Guo

With the rapid development of cloud-based services, large language models have become increasingly accessible through various web platforms. However, this accessibility has also led to growing risks of model abuse. LLM watermarking has emerged as an effective approach to mitigate such misuse and protect intellectual property. Existing watermarking algorithms, however, primarily focus on defending against paraphrase attacks while overlooking piggyback spoofing attacks, which can inject harmful content, compromise watermark reliability, and undermine trust in attribution. To address this limitation, we propose DualGuard, the first watermarking algorithm capable of defending against both paraphrase and spoofing attacks. DualGuard employs the adaptive dual-stream watermarking mechanism, in which two complementary watermark signals are dynamically injected based on the semantic content. This design enables DualGuard not only to detect but also to trace spoofing attacks, thereby ensuring reliable and trustworthy watermark detection. Extensive experiments conducted across multiple datasets and language models demonstrate that DualGuard achieves excellent detectability, robustness, traceability, and text quality, effectively advancing the state of LLM watermarking for real-world applications.

Xu Shen, Qi Zhang, Song Wang, Zhen Tan, Xinyu Zhao, Laura Yao, Vaishnav Tadiparthi, Hossein Nourkhiz Mahjoub, Ehsan Moradi Pari, Kwonjoon Lee 等

Large Language Model based multi-agent systems (MAS) excel at collaborative problem solving but remain brittle to cascading errors: a single faulty step can propagate across agents and disrupt the trajectory. In this paper, we present MASC, a metacognitive framework that endows MAS with real-time, unsupervised, step-level error detection and self-correction. MASC rethinks detection as history-conditioned anomaly scoring via two complementary designs: (1) Next-Execution Reconstruction, which predicts the embedding of the next step from the query and interaction history to capture causal consistency, and (2) Prototype-Guided Enhancement, which learns a prototype prior over normal-step embeddings and uses it to stabilize reconstruction and anomaly scoring under sparse context (e.g., early steps). When an anomaly step is flagged, MASC triggers a correction agent to revise the acting agent’s output before information flows downstream. On the Who When benchmark, MASC consistently outperforms all baselines, achieving up to 7.8% AUC-ROC improvement in the challenging w/o GT setting, and further delivers consistent gains on AgentErrorBench. When plugged into diverse MAS frameworks, it delivers consistent end-to-end gains across architectures, confirming that our metacognitive monitoring and targeted correction can mitigate error propagation with minimal overhead.

Chao Xue, Yao Wang, Mengqiao Liu, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Chenyao Lu, Lei Jiang, Yu Lu 等

Recent advancements in the Generative Reward Model (GRM) have demonstrated its potential to enhance the reasoning abilities of LLMs through Chain-of-Thought (CoT) prompting. Despite these gains, existing implementations of GRM suffer from two critical limitations. First, CoT prompting is applied indiscriminately to all inputs regardless of their inherent complexity. This introduces unnecessary computational costs for tasks amenable to fast, direct inference. Second, existing approaches primarily rely on voting-based mechanisms to evaluate CoT outputs, which often lack granularity and precision in assessing reasoning quality. In this paper, we propose E-GRM, an efficient generative reward modeling framework grounded in model-internal uncertainty. E-GRM leverages the convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed, without relying on handcrafted features or task-dependent signals. To improve reward fidelity, we introduce a lightweight discriminative scorer trained with a hybrid regression–ranking objective to provide fine-grained evaluation of reasoning paths. Experiments on multiple reasoning benchmarks show that E-GRM substantially reduces inference cost while consistently improving answer accuracy, demonstrating that model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling.

Ruoran Li, Xinghua Zhang, Haiyang Yu, Shitong Duan, Xiang Li, Wenxin Xiang, Chonghua Liao, Xudong Guo, Yongbin Li, Jinli Suo

Long-horizon agents face the challenge of growing context size during interaction with environment, which degrades the performance and stability. Existing methods typically introduce the external memory module and look up the relevant information from the stored memory, which prevents the model itself from proactively managing its memory content and aligning with the agent’s overarching task objectives. To address these limitations, we propose the self-memory policy optimization algorithm (MemPO), which enables the agent (policy model) to autonomously summarize and manage their memory during interaction with environment. By improving the credit assignment mechanism based on memory effectiveness, the policy model can selectively retain crucial information, significantly reducing token consumption while preserving task performance. Extensive experiments and analyses confirm that MemPO achieves absolute F1 score gains of 25.98 over the base model and 7.1 over the previous SOTA baseline, while reducing token usage by 67.58% and 73.12%.

Chenxi Zhou, Pengfei Cao, Jiang Li, Bohan Yu, Jinyu Ye, Jun Zhao, Kang Liu

Post-Training Quantization (PTQ) is a critical strategy for efficient large language models (LLMs) deployment. However, existing scaling laws primarily focus on general performance, overlooking crucial fine-grained factors and how quantization differentially impacts diverse knowledge capabilities. To address this, we establish Task-Stratified Knowledge Scaling Laws. By stratifying capabilities into memorization, application, and reasoning, we develop a framework that unifies model size, bit-width, and fine-grained factors: group size and calibration set size. Validated on 293 diverse PTQ configurations, our framework demonstrates strong fit and cross-architecture consistency. It reveals distinct sensitivities across knowledge capabilities: reasoning is precision-critical, application is scale-responsive, and memorization is calibration-sensitive. We highlight that in low-bit scenarios, optimizing these fine-grained factors is essential for preventing performance collapse. These findings provide an empirically-backed foundation for designing knowledge-aware quantization strategies.

Zhou Ziheng, Jiakun Ding, Zhaowei Zhang, Ruosen Gao, Ying Nian Wu, Demetri Terzopoulos, Yipeng Kang, Fangwei Zhong, Junqi Wang

Principle-based alignment often lacks context sensitivity and completeness. Grounded in Theory of Mind, we propose role conditioning as a compact alternative: social roles (e.g., mother, judge) implicitly encode both values and the cognitive schemas required to apply them. We introduce a training-free pipeline featuring a role-conditioned generator and iterative role-based critics for refinement. Across five model families, our approach consistently outperforms principle-based, Chain-of-Thought (CoT) and other baselines across benchmarks. Notably, it reduces unsafe outputs on the WildJailbreak benchmark from 81.4% to 3.6% with DeepSeek-V3. Not only for common safety benchmarks, it consistently applies for agentic safety tasks. These results establish role assignment as a powerful, interpretable paradigm for AI alignment and LLM-as-a-Judge construction.

Ryo Kishino, Yusuke Takase, Momose Oyama, Hiroaki Yamagiwa, Hidetoshi Shimodaira

Log-likelihood vectors define a common space for comparing language models as probability distributions, enabling unified comparisons across heterogeneous settings. We extend this framework to training checkpoints and intermediate layers, and establish a consistent scale for KL divergence across pretraining, model size, random seeds, quantization, fine-tuning, and layers. Analysis of Pythia pretraining trajectories further shows that changes in log-likelihood space, as measured by the scaling behavior of KL divergence, are much smaller than in weight space, resulting in subdiffusive learning trajectories and early stabilization of language-model behavior despite weight drift.

Chenxi Zhou, Pengfei Cao, Jiang Li, Bohan Yu, Jinyu Ye, Jun Zhao, Kang Liu

Post-Training Quantization (PTQ) is critical for the efficient deployment of Large Language Models (LLMs). While 4-bit quantization is widely regarded as an optimal trade-off, reducing the precision to 2-bit usually triggers a catastrophic “performance cliff.” It remains unclear whether the underlying mechanisms differ fundamentally. Consequently, we conduct a systematic mechanistic analysis, revealing two qualitatively distinct failure modes: Signal Degradation, where the computational patterns remain intact but information precision is impaired by cumulative error; and Computation Collapse, where key components fail to function, preventing correct information processing and destroying the signal in the early layers. Guided by this diagnosis, we conduct mechanism-aware interventions, demonstrating that targeted, training-free repair can mitigate Signal Degradation, but remains ineffective for Computation Collapse. Our findings provide a systematic diagnostic framework for PTQ failures and suggest that addressing Computation Collapse requires structural reconstruction rather than mere compensation.

Xiaoyong Mei, Tingting Zuo, Da Chen, Guangyu Hu, Xiangyu Wen, Chao Duan, Mingyan Zhang, Fudan Zheng

Multi-role dialogue summarization requires modeling complex interactions among multiple speakers while preserving role-specific information and factual consistency. However, most existing methods optimize for automatic metrics such as ROUGE and BERTScore, which favor surface-level imitation of references rather than genuine gains in faithfulness or alignment with human preferences. We propose a novel framework that couples explicit cognitive-style reasoning with reward-based optimization for multi-role dialogue summarization. Our method first distills structured reasoning traces (e.g., step-by-step inferences and intermediate reflections) from a large teacher model and uses them as auxiliary supervision to initialize a reasoning-aware summarizer via staged supervised fine-tuning. It then applies GRPO with a dual-principle reward that blends metric-based signals with human-aligned criteria targeting key information coverage, implicit inference, factual faithfulness, and conciseness. Experiments on multilingual multi-role dialogue benchmarks show that our method matches strong baselines on ROUGE and BERTScore. Specifically, results on CSDS confirm the framework’s stability in semantic consistency, while in-depth analysis on SAMSum demonstrates clear gains in factual faithfulness and model-based preference alignment. These findings underscore the value of reasoning-aware and preference-aware training for reliable dialogue summarization. Code will be made accessible upon acceptance, checkpoints and datasets are now available at https://huggingface.co/NebulaPixel.

Yifei Gong, Xing Wu, Wenda Liu, Tukang

Computer-Aided Design (CAD) is an expert-level task that relies on long-horizon reasoning and coherent modeling actions. Large Language Models (LLMs) have shown remarkable advancements in enabling language agents to tackle real-world tasks. Notably, there has been no investigation into how tool-using LLMs optimally interact with CAD engines, hindering the emergence of LLM-based agentic text-to-CAD modeling systems. We propose ToolCAD, a novel agentic CAD framework deploying LLMs as tool-using agents for text-to-CAD generation. Furthermore, we introduce an interactive CAD modeling gym to rollout reasoning and tool-augmented interaction trajectories with the CAD engine, incorporating hybrid feedback and human supervision. Meanwhile, an end-to-end post-training strategy is presented to enable the LLM agent to elicit refined CAD Modeling Chain of Thought (CAD-CoT) and evolve into proficient CAD tool-using agents via online curriculum reinforcement learning. Our findings demonstrate ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for more accessible and robust autonomous text-to-CAD modeling systems.

Zhuoqun Li, Zhaopei Huang, Wenxuan Wang, Qin Jin

Understanding multimodal user personas in long-term dialogues is essential for building personalized and human-like dialogue systems. However, existing datasets suffer from limited persona diversity and static, overly simplified settings, making them insufficient for capturing the complexity of real-world interactions. To address these limitations, we introduce LongMP-Bench, a benchmark designed to evaluate the capabilities of models in understanding evolving user personas within long-term multimodal dialogues. We present a multi-step, scalable data construction pipeline that generates long-term interaction records centered around multimodal personas, followed by human refinement for quality assurance. The resulting dataset contains long conversations from 150 users, each exhibiting visual consistency and dynamic persona development over time. Built on this dataset, we define a suite of tasks to comprehensively assess models’ ability to track persona evolution, integrate visual and textual inputs, and apply persona understanding in realistic dialogue scenarios. Extensive experiments on LongMP-Bench highlight the substantial challenges in multimodal persona understanding, especially in tracking persona shifts and leveraging multimodal context effectively. We will release our benchmark and code to facilitate future research in multimodal and personalized dialogue systems.

Zhe Chen, Jiaao Yu, Honglin Li

Length-controllable text generation (LCTG) is essential for tasks like text summarization and report generation. However, large language models (LLMs) have limited awareness of output length, so precise control over the length of generated text remains a significant challenge. Most existing methods focus on prompt-based frameworks, position encoding, and reinforcement learning for model training. These approaches may affect semantic quality, and struggle to maintain consistent length control across different models and tasks. In this paper, we propose DcLM, a model-agnostic approach that introduces dynamic length markers to guide length-controllable outputs. During training, the model leverages these markers as in-context information, without learning to generate them. At inference time, an external word counter and injected length information guide the model to produce outputs of accurate lengths. We evaluate our method across multiple datasets, and the experimental results demonstrate that DcLM significantly reduces length deviation, showcasing its robust generalization ability across various length scales and tasks.