Le Sun

Yanjiang Liu, Jie Lou, Xinyan Guan et al.

On-policy distillation transfers reasoning capabilities by training a student model on its own generated trajectories using token-level feedback from a teacher. However, we identify a critical bottleneck, \textbf{Supervision Fidelity Decay (SFD)}: as student-generated prefixes lengthen, the teacher's next-token distribution becomes less confident and less discriminative. Consequently, the teacher-dependent corrective signal in reverse-KL distillation weakens, causing student drift to compound across long reasoning chains. To mitigate SFD, we introduce \textbf{Lookahead Group Reward (\ours{})}. Building on the insight that next-step teacher confidence reflects the discriminative strength of future reverse-KL supervision, \ours{} evaluates the student's top-K candidate tokens by the teacher confidence they induce at the subsequent step and assigns a group-normalized reward. To maintain computational efficiency, we further design an entropy-triggered tree-attention mechanism. Across six math and code benchmarks, \ours{} improves mean@8 by \textbf{2.57} points over OPD for a 7B student, with gains increasing in longer-generation and reaching +\textbf{4.92} points on AIME-26 at 39k tokens.

Xinyu Lu, Tianshu Wang, Pengbo Wang et al.

Current AI benchmarks evaluate agents on task execution within human-designed workflows. These evaluations fundamentally fail to measure a critical next-level capability: whether models can autonomously develop agent systems. We introduce the Meta-Agent Challenge (MAC), an evaluation framework designed to test the capacity of frontier models for autonomous agent development. Specifically, a code agent (the meta-agent) is given a sandboxed environment, an evaluation API, and a time limitation to iteratively program an agent artifact that maximizes performance on a held-out test set across five domains. To ensure evaluation integrity, this framework is secured by multi-layer defenses against reward hacking. Leveraging this framework, we demonstrate that meta-agents rarely match human-engineered baseline policies, and the few that do are dominated by proprietary frontier models. Moreover, the design process exhibits high variance, and high optimization pressure surfaces emergent adversarial behaviors like ground-truth exfiltration-highlighting critical deficits in both robustness and model alignment. Ultimately, MAC provides a rigorous, open-source benchmark for autonomous AI research and development, offering an empirical proxy for evaluating recursive self-improvement. Benchmark is publicly available at: https://github.com/ant-research/meta-agent-challenge.

Xiaoxuan Peng, Kaiqi Zhang, Xinyu Lu et al.

Mastering terminal environments requires language agents capable of multi-step planning, feedback-grounded execution, and dynamic state adaptation. However, training such agents is currently bottlenecked by a reliance on scraped external repositories, which limits domain diversity, environment controllability, and the targeting of specific capability deficits. We introduce LiteCoder-Terminal-Gen, a zero-dependency synthesis pipeline that autonomously generates executable and verifiable terminal training environments directly from domain specifications. Using this framework, we construct two large-scale resources: LiteCoder-Terminal-SFT, comprising 11,255 expert trajectories across 10 domains, and LiteCoder-Terminal-RL, featuring 602 verifiable environments for trajectory-level preference optimization. Supervised fine-tuning of Qwen-family models on our SFT dataset yields agents that significantly outperform their base counterparts. Notably, our 32B variant achieves 29.06%, 18.54%, and 34.00% pass@1 on Terminal Bench 1.0, 2.0, and Pro, respectively. Furthermore, applying Direct Multi-turn Preference Optimization (DMPO) on our RL environments yields additional performance gains. These results systematically demonstrate that fully synthetic, executable environments offer a scalable and verifiable supervision signal for mastering complex, real-world command-line workflows.

Boxi Cao, Hongyu Lin, Xianpei Han et al.

Prompt-based probing has been widely used in evaluating the abilities of pretrained language models (PLMs). Unfortunately, recent studies have discovered such an evaluation may be inaccurate, inconsistent and unreliable. Furthermore, the lack of understanding its inner workings, combined with its wide applicability, has the potential to lead to unforeseen risks for evaluating and applying PLMs in real-world applications. To discover, understand and quantify the risks, this paper investigates the prompt-based probing from a causal view, highlights three critical biases which could induce biased results and conclusions, and proposes to conduct debiasing via causal intervention. This paper provides valuable insights for the design of unbiased datasets, better probing frameworks and more reliable evaluations of pretrained language models. Furthermore, our conclusions also echo that we need to rethink the criteria for identifying better pretrained language models. We openly released the source code and data at https://github.com/c-box/causalEval.

Qiming Zhu, Jialun Cao, Xuanang Chen et al.

Current research on large language models (LLMs) with retrieval-augmented code generation (RACG) has largely focused on single-language settings, leaving their cross-lingual effectiveness underexplored. Multilingual RACG systems are increasingly important for migrating and reusing code across programming languages (PLs), a common yet challenging task in modern software development. To systematically study cross-lingual code knowledge transfer in RACG, we construct a dataset covering 13 PLs with nearly 14K instances. Our experiments reveal three key insights: (1) Knowledge transfer in RACG across PLs is non-trivial even using direct injection. (2) RACG exhibits unequal cross-lingual knowledge transfer, and its efficacy depends on linguistic affinity of PL pair and diversity of LLM pretraining corpus. (3) RACG shows limited reliance on natural language information embedded in code when equipped with a code-specific retriever. These findings provide practical guidance for designing effective multilingual RACG systems. https://github.com/icip-cas/Cross-Lingual-RACG

Shu Chen, Xinyan Guan, Yaojie Lu et al.

Manually annotating instruction data for large language models is difficult, costly, and hard to scale. Meanwhile, current automatic annotation methods typically rely on distilling synthetic data from proprietary LLMs, which not only limits the upper bound of the quality of the instruction data but also raises potential copyright issues. In this paper, we propose REInstruct, a simple and scalable method to automatically build instruction data from an unlabeled corpus without heavy reliance on proprietary LLMs and human annotation. Specifically, REInstruct first selects a subset of unlabeled texts that potentially contain well-structured helpful and insightful content and then generates instructions for these texts. To generate accurate and relevant responses for effective and robust training, REInstruct further proposes a rewriting-based approach to improve the quality of the generated instruction data. By training Llama-7b on a combination of 3k seed data and 32k synthetic data from REInstruct, fine-tuned model achieves a 65.41\% win rate on AlpacaEval leaderboard against text-davinci-003, outperforming other open-source, non-distilled instruction data construction methods. The code is publicly available at \url{https://github.com/cs32963/REInstruct}.

Xinru Yan, Boxi Cao, Yaojie Lu et al.

Native Omni-modal Large Language Models (OLLMs) have shifted from pipeline architectures to unified representation spaces. However, this native integration gives rise to a critical yet underexplored phenomenon: modality preference. To bridge this gap, we first systematically quantify modality preference of OLLMs using a newly-curated conflict-based benchmark and the modality selection rate metric. Our evaluation of ten representative OLLMs reveals a notable paradigm shift: unlike the ``text-dominance'' of traditional VLMs, most OLLMs exhibit a pronounced visual preference. To further understand the underlying mechanism, we conduct layer-wise probing and demonstrate that such modality preference is not static but emerges progressively in the mid-to-late layers. Building upon these insights, we leverage these internal signals to diagnose cross-modal hallucinations, achieving competitive performance across three downstream multi-modal benchmarks without task-specific data. Our work provides both a mechanistic understanding and a practical tool for building more trustworthy OLLMs. Our code and related resources are publicly available at: https://github.com/icip-cas/OmniPreference

Yaojie Lu, Qing Liu, Dai Dai et al.

Information extraction suffers from its varying targets, heterogeneous structures, and demand-specific schemas. In this paper, we propose a unified text-to-structure generation framework, namely UIE, which can universally model different IE tasks, adaptively generate targeted structures, and collaboratively learn general IE abilities from different knowledge sources. Specifically, UIE uniformly encodes different extraction structures via a structured extraction language, adaptively generates target extractions via a schema-based prompt mechanism - structural schema instructor, and captures the common IE abilities via a large-scale pre-trained text-to-structure model. Experiments show that UIE achieved the state-of-the-art performance on 4 IE tasks, 13 datasets, and on all supervised, low-resource, and few-shot settings for a wide range of entity, relation, event and sentiment extraction tasks and their unification. These results verified the effectiveness, universality, and transferability of UIE.

Jie Lou, Yaojie Lu, Dai Dai et al.

The challenge of information extraction (IE) lies in the diversity of label schemas and the heterogeneity of structures. Traditional methods require task-specific model design and rely heavily on expensive supervision, making them difficult to generalize to new schemas. In this paper, we decouple IE into two basic abilities, structuring and conceptualizing, which are shared by different tasks and schemas. Based on this paradigm, we propose to universally model various IE tasks with Unified Semantic Matching (USM) framework, which introduces three unified token linking operations to model the abilities of structuring and conceptualizing. In this way, USM can jointly encode schema and input text, uniformly extract substructures in parallel, and controllably decode target structures on demand. Empirical evaluation on 4 IE tasks shows that the proposed method achieves state-of-the-art performance under the supervised experiments and shows strong generalization ability in zero/few-shot transfer settings.

Qiaoyu Tang, Ziliang Deng, Hongyu Lin et al.

Enabling large language models to utilize real-world tools effectively is crucial for achieving embodied intelligence. Existing approaches to tool learning have either primarily relied on extremely large language models, such as GPT-4, to attain generalized tool-use abilities in a zero-shot manner, or utilized supervised learning to train limited scopes of tools on compact models. However, it remains uncertain whether smaller language models can achieve generalized tool-use abilities without tool-specific training. To address this question, this paper introduces ToolAlpaca, a novel framework designed to automatically generate a diverse tool-use corpus and learn generalized tool-use abilities on compact language models with minimal human intervention. Specifically, ToolAlpaca first automatically creates a highly diversified tool-use corpus by building a multi-agent simulation environment. The corpus contains 3938 tool-use instances from more than 400 real-world tool APIs spanning 50 distinct categories. Subsequently, the constructed corpus is employed to fine-tune compact language models, resulting in two models, namely ToolAlpaca-7B and ToolAlpaca-13B, respectively. Finally, we evaluate the ability of these models to utilize previously unseen tools without specific training. Experimental results demonstrate that ToolAlpaca achieves effective generalized tool-use capabilities comparable to those of extremely large language models like GPT-3.5, demonstrating that learning generalized tool-use ability is feasible for compact language models.

Hao Zheng, Guozhao Mo, Xinru Yan et al.

Presentation generation requires deep content research, coherent visual design, and iterative refinement based on observation. However, existing presentation agents often rely on predefined workflows and fixed templates. To address this, we present DeepPresenter, an agentic framework that adapts to diverse user intents, enables effective feedback-driven refinement, and generalizes beyond a scripted pipeline. Specifically, DeepPresenter autonomously plans, renders, and revises intermediate slide artifacts to support long-horizon refinement with environmental observations. Furthermore, rather than relying on self-reflection over internal signals (e.g., reasoning traces), our environment-grounded reflection conditions the generation process on perceptual artifact states (e.g., rendered slides), enabling the system to identify and correct presentation-specific issues during execution. Results on the evaluation set covering diverse presentation-generation scenarios show that DeepPresenter achieves state-of-the-art performance, and the fine-tuned 9B model remains highly competitive at substantially lower cost. Our project is available at: https://github.com/icip-cas/PPTAgent

Jiasheng Zheng, Boxi Cao, Boxi Yu et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as the cornerstone for shaping the remarkable coding abilities of Large Language Models (LLMs). However, the scalability of RLVR is severely constrained by the scarcity of sufficiently challenging verifiable code tasks that target near the model's edge of competence. Prior studies often rely on heuristic seed expansions for data synthesis, which severely limits both novelty and difficulty. Consequently, the training value of such data fails to scale proportionally with the size of its synthesis. To this end, we propose Atomic Decomposition and Recombination (ADR), a novel framework that generates verifiable code tasks via decomposition into atomic elements and controlled recombination, thereby enabling the generation of genuinely novel and challenging verifiable code tasks. Experiments and analysis demonstrate that ADR achieves superior originality, difficulty, diversity, and test quality over existing baselines, and consistently delivers greater improvements in code ability across RLVR in diverse downstream domains, including algorithmic programming, tool usage, and data science. Our work sheds light on a new paradigm for novel code task synthesis and scalable RLVR training.

Xinyan Guan, Yanjiang Liu, Hongyu Lin et al.

Incorporating factual knowledge in knowledge graph is regarded as a promising approach for mitigating the hallucination of large language models (LLMs). Existing methods usually only use the user's input to query the knowledge graph, thus failing to address the factual hallucination generated by LLMs during its reasoning process. To address this problem, this paper proposes Knowledge Graph-based Retrofitting (KGR), a new framework that incorporates LLMs with KGs to mitigate factual hallucination during the reasoning process by retrofitting the initial draft responses of LLMs based on the factual knowledge stored in KGs. Specifically, KGR leverages LLMs to extract, select, validate, and retrofit factual statements within the model-generated responses, which enables an autonomous knowledge verifying and refining procedure without any additional manual efforts. Experiments show that KGR can significantly improve the performance of LLMs on factual QA benchmarks especially when involving complex reasoning processes, which demonstrates the necessity and effectiveness of KGR in mitigating hallucination and enhancing the reliability of LLMs.

Qianhao Yuan, Jie Lou, Zichao Li et al.

Typical search agents concatenate the entire interaction history into the LLM context, preserving information integrity but producing long, noisy contexts, resulting in high computation and memory costs. In contrast, using only the current turn avoids this overhead but discards essential information. This trade-off limits the scalability of search agents. To address this challenge, we propose MemSearcher, an agent workflow that iteratively maintains a compact memory and combines the current turn with it. At each turn, MemSearcher fuses the user's question with the memory to generate reasoning traces, perform search actions, and update memory to retain only information essential for solving the task. This design stabilizes context length across multi-turn interactions, improving efficiency without sacrificing accuracy. To optimize this workflow, we introduce multi-context GRPO, an end-to-end RL framework that jointly optimize reasoning, search strategies, and memory management of MemSearcher Agents. Specifically, multi-context GRPO samples groups of trajectories under different contexts and propagates trajectory-level advantages across all conversations within them. Trained on the same dataset as Search-R1, MemSearcher achieves significant improvements over strong baselines on seven public benchmarks: +11% on Qwen2.5-3B-Instruct and +12% on Qwen2.5-7B-Instruct relative average gains. Notably, the 3B-based MemSearcher even outperforms 7B-based baselines, demonstrating that striking a balance between information integrity and efficiency yields both higher accuracy and lower computational overhead. The code and models will be publicly available at https://github.com/icip-cas/MemSearcher

Ning Bian, Xianpei Han, Le Sun et al.

Large language models (LLMs) have made significant progress in NLP. However, their ability to memorize, represent, and leverage commonsense knowledge has been a well-known pain point. In this paper, we specifically focus on ChatGPT, a widely used and easily accessible LLM, and ask the following questions: (1) Can ChatGPT effectively answer commonsense questions? (2) Is ChatGPT aware of the underlying commonsense knowledge for answering a specific question? (3) Is ChatGPT knowledgeable in commonsense? (4) Can ChatGPT effectively leverage commonsense for answering questions? We conduct a series of experiments on 11 datasets to evaluate ChatGPT's commonsense abilities, including answering commonsense questions, identifying necessary knowledge, generating knowledge descriptions, and using knowledge descriptions to answer questions again. Experimental results show that: (1) ChatGPT can achieve good QA accuracies in commonsense tasks, while still struggling with certain domains of datasets. (2) ChatGPT is knowledgeable, and can accurately generate most of the commonsense knowledge using knowledge prompts. (3) Despite its knowledge, ChatGPT is an inexperienced commonsense problem solver, which cannot precisely identify the needed commonsense for answering a specific question. These findings raise the need to explore improved mechanisms for effectively incorporating commonsense into LLMs like ChatGPT, such as better instruction following and commonsense guidance.

Qiming Zhu, Jialun Cao, Yaojie Lu et al.

Code benchmarks such as HumanEval are widely adopted to evaluate the capabilities of Large Language Models (LLMs), providing insights into their strengths and weaknesses. However, current benchmarks primarily exercise LLMs' capability on common coding tasks (e.g., bubble sort, greatest common divisor), leaving domain-specific coding tasks (e.g., computation, system, cryptography) unexplored. To fill this gap, we propose a multi-domain code benchmark, DOMAINEVAL, designed to evaluate LLMs' coding capabilities thoroughly. Our pipeline works in a fully automated manner, enabling a push-bottom construction from code repositories into formatted subjects under study. Interesting findings are observed by evaluating 12 representative LLMs against DOMAINEVAL. We notice that LLMs are generally good at computation tasks while falling short on cryptography and system coding tasks. The performance gap can be as much as 68.94% (80.94% - 12.0%) in some LLMs. We also observe that generating more samples can increase the overall performance of LLMs, while the domain bias may even increase. The contributions of this study include a code generation benchmark dataset DOMAINEVAL, encompassing six popular domains, a fully automated pipeline for constructing code benchmarks, and an identification of the limitations of LLMs in code generation tasks based on their performance on DOMAINEVAL, providing directions for future research improvements. The leaderboard is available at https://domaineval.github.io/.

Dong Xu, Jialun Cao, Guozhao Mo et al.

Formal specification is essential for rigorous program verification, yet writing correct specifications remains costly and difficult to automate. Although large language models (LLMs) and agents have shown promising progress, their true capabilities and failure modes remain unclear. We present the first systematic and contamination-aware study of LLM- and agent-based formal specification generation for C programs. We introduce LiveFMBench, a continuously evolving benchmark of 630 ACSL (ANSI/ISO C Specification Language)-annotated C programs, including 360 newly collected cases designed to mitigate data leakage. Using this benchmark, we evaluate direct prompting with different sampling sizes, reasoning-enabled (thinking mode) inference, the agentic pipeline, and perform a fine-grained failure analysis. Experimental results reveal that naive evaluation substantially overestimates performance because models under direct prompting may exhibit unfaithful behaviors, such as deceiving automated provers or ignoring code-context constraints; after excluding such cases, the true specification generation accuracy drops by approximately 20\%. We further find that both increased sampling and thinking mode significantly improve success rates, with smaller models benefiting more from thinking mode. Agentic pipelines are particularly effective under low sampling budgets and on harder datasets. Failure analysis further shows that incorrect loop invariants are the dominant error type, while agentic pipelines notably reduce assertion errors. These results expose fundamental limitations in current LLM-based approaches and suggest they remain far from replacing human-authored formal specifications. We release LiveFMBench at https://huggingface.co/datasets/fm-universe/Live-FM-Bench and all evaluation artifacts to support future research.

Ruiyang Xu, Jialun Cao, Yaojie Lu et al.

Code benchmarks such as HumanEval are widely adopted to evaluate Large Language Models' (LLMs) coding capabilities. However, there is an unignorable programming language bias in existing code benchmarks -- over 95% code generation benchmarks are dominated by Python, leaving the LLMs' capabilities in other programming languages such as Java and C/C++ unknown. Moreover, coding task bias is also crucial. Most benchmarks focus on code generation capability, while benchmarks for code reasoning (given input, reasoning output; and given output, reasoning input), an essential coding capability, are insufficient. Yet, constructing multi-lingual benchmarks can be expensive and labor-intensive, and codes in contest websites such as Leetcode suffer from data contamination during training. To fill this gap, we propose CRUXEVAL-X, a multi-lingual code reasoning benchmark that contains 19 programming languages. It comprises at least 600 subjects for each language, along with 19K content-consistent tests in total. In particular, the construction pipeline of CRUXEVAL-X works in a fully automated and test-guided manner, which iteratively generates and repairs based on execution feedback. Also, to cross language barriers (e.g., dynamic/static type systems in Python/C++), we formulated various transition rules between language pairs to facilitate translation. Our intensive evaluation of 24 representative LLMs reveals the correlation between language pairs. For example, TypeScript and JavaScript show a significant positive correlation, while Racket has less correlation with other languages. More interestingly, even a model trained solely on Python can achieve at most 34.4% Pass@1 in other languages, revealing the cross-language generalization of LLMs.

Jiawei Chen, Qing Liu, Hongyu Lin et al.

Few-shot NER needs to effectively capture information from limited instances and transfer useful knowledge from external resources. In this paper, we propose a self-describing mechanism for few-shot NER, which can effectively leverage illustrative instances and precisely transfer knowledge from external resources by describing both entity types and mentions using a universal concept set. Specifically, we design Self-describing Networks (SDNet), a Seq2Seq generation model which can universally describe mentions using concepts, automatically map novel entity types to concepts, and adaptively recognize entities on-demand. We pre-train SDNet with large-scale corpus, and conduct experiments on 8 benchmarks from different domains. Experiments show that SDNet achieves competitive performances on all benchmarks and achieves the new state-of-the-art on 6 benchmarks, which demonstrates its effectiveness and robustness.

Jialong Tang, Hongyu Lin, Meng Liao et al. · tencent-ai

Procedural text understanding requires machines to reason about entity states within the dynamical narratives. Current procedural text understanding approaches are commonly \textbf{entity-wise}, which separately track each entity and independently predict different states of each entity. Such an entity-wise paradigm does not consider the interaction between entities and their states. In this paper, we propose a new \textbf{scene-wise} paradigm for procedural text understanding, which jointly tracks states of all entities in a scene-by-scene manner. Based on this paradigm, we propose \textbf{S}cene \textbf{G}raph \textbf{R}easoner (\textbf{SGR}), which introduces a series of dynamically evolving scene graphs to jointly formulate the evolution of entities, states and their associations throughout the narrative. In this way, the deep interactions between all entities and states can be jointly captured and simultaneously derived from scene graphs. Experiments show that SGR not only achieves the new state-of-the-art performance but also significantly accelerates the speed of reasoning.

Qianhao Yuan, Jie Lou, Xing Yu et al.

Multimodal Large Language Models (MLLMs) still struggle with fine-grained visual understanding, where answers often depend on small but decisive evidence in the full image. We observe a regional-to-global perception gap: the same MLLM answers fine-grained questions more accurately when conditioned on evidence-centered crops than on the corresponding full images, suggesting that many failures stem from difficulty to focus on relevant evidence rather than insufficient local recognition ability. Motivated by this observation, we propose Vision-OPD (Vision On-Policy Distillation), a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy. Vision-OPD instantiates two conditional policies from the same MLLM: a crop-conditioned teacher and a full-image-conditioned student. The student generates on-policy rollouts, and Vision-OPD minimizes token-level divergence between the teacher and student next-token distributions along these rollouts. This enables the model to internalize the benefit of visual zooming without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use. Experiments on multiple fine-grained visual understanding benchmarks show that Vision-OPD models achieve competitive or superior performance against much larger open-source, closed-source, and "Thinking-with-Images" agentic models.

Fangchao Liu, Hongyu Lin, Xianpei Han et al.

Low-shot relation extraction~(RE) aims to recognize novel relations with very few or even no samples, which is critical in real scenario application. Few-shot and zero-shot RE are two representative low-shot RE tasks, which seem to be with similar target but require totally different underlying abilities. In this paper, we propose Multi-Choice Matching Networks to unify low-shot relation extraction. To fill in the gap between zero-shot and few-shot RE, we propose the triplet-paraphrase meta-training, which leverages triplet paraphrase to pre-train zero-shot label matching ability and uses meta-learning paradigm to learn few-shot instance summarizing ability. Experimental results on three different low-shot RE tasks show that the proposed method outperforms strong baselines by a large margin, and achieve the best performance on few-shot RE leaderboard.

Boxi Cao, Hongyu Lin, Xianpei Han et al.

Knowledge plays a critical role in artificial intelligence. Recently, the extensive success of pre-trained language models (PLMs) has raised significant attention about how knowledge can be acquired, maintained, updated and used by language models. Despite the enormous amount of related studies, there still lacks a unified view of how knowledge circulates within language models throughout the learning, tuning, and application processes, which may prevent us from further understanding the connections between current progress or realizing existing limitations. In this survey, we revisit PLMs as knowledge-based systems by dividing the life circle of knowledge in PLMs into five critical periods, and investigating how knowledge circulates when it is built, maintained and used. To this end, we systematically review existing studies of each period of the knowledge life cycle, summarize the main challenges and current limitations, and discuss future directions.

Xin Zheng, Jie Lou, Boxi Cao et al.

Self-critic has become a crucial mechanism for enhancing the reasoning performance of LLMs. However, current approaches mainly involve basic prompts for intuitive instance-level feedback, which resembles System-1 processes and limits the reasoning capabilities. Moreover, there is a lack of in-depth investigations into the relationship between LLM's ability to criticize and its task-solving performance. To address these issues, we propose Critic-CoT, a novel framework that pushes LLMs toward System-2-like critic capability. Through a step-wise CoT reasoning paradigm and the automatic construction of distant-supervision data without human annotation, Critic-CoT enables LLMs to engage in slow, analytic self-critique and refinement, thereby improving their reasoning abilities. Experiments on GSM8K and MATH demonstrate that our enhanced model significantly boosts task-solving performance by filtering out invalid solutions or iterative refinement. Furthermore, we investigate the intrinsic correlation between critique and task-solving abilities within LLMs, discovering that these abilities can mutually reinforce each other rather than conflict.

Jiasheng Zheng, Boxi Cao, Zhengzhao Ma et al.

In recent years, researchers have proposed numerous benchmarks to evaluate the impressive coding capabilities of large language models (LLMs). However, current benchmarks primarily assess the accuracy of LLM-generated code, while neglecting other critical dimensions that also significantly impact code quality in real-world development. Moreover, relying exclusively on correctness as the guiding metric renders LLMs susceptible to data contamination. Therefore, this paper proposes the RACE benchmark, which comprehensively evaluates the quality of code generated by LLMs across 4 dimensions: Readability, mAintainability, Correctness, and Efficiency. Specifically, considering the demand-dependent nature of dimensions beyond correctness, we design various types of user requirements for each dimension to assess the model's ability to generate correct code that also meets user demands. We analyze 28 representative LLMs based on RACE and find that: 1) current correctness-centric benchmarks fail to capture the multifaceted requirements of code in real-world scenarios, while RACE provides a comprehensive evaluation that reveals the defects of LLMs across multiple dimensions; 2) the RACE benchmark serves as an effective tool for resisting the risk of data contamination; 3) even the most advanced code LLMs still encounter significant challenges in customized requirements involving complex instructions; 4) most LLMs exhibit an inherent preference for specific coding style. These findings highlight the need for a multidimensional evaluation of code LLMs, emphasizing metrics beyond correctness for real-world applications. Future efforts should aim to develop novel learning algorithms to enhance code generation under varied constraints and improve coverage and usability for diverse user needs.

Zhuoqun Li, Boxi Cao, Guiping Jiang et al.

Metaphorical videos are prevalent across various real-world scenarios to convey complex ideas, and understanding them typically requires high-order cognitive capabilities. The lack of systematic studies on metaphorical video understanding not only constrains the real-world applicability of MLLMs but also impedes the thorough assessment of their high-order cognitive capabilities. To bridge this gap, we propose MetaphorVU-Bench, the first systematic and comprehensive benchmark dedicated to metaphorical video understanding. Through experiments, we find current MLLMs struggle with accurate metaphorical video understanding, lagging far behind human level, primarily due to defective cross-domain mapping. Motivated by this finding, we construct a metaphor knowledge graph as mapping augmentation and propose MetaphorBoost, an inference-time enhancement framework achieving consistent performance improvement. Our benchmark, analysis, and method provide useful insights and a foundation for future research on advancing MLLMs.

Zichao Li, Aizier Abulaiti, Yaojie Lu et al.

Document Structured Extraction (DSE) aims to extract structured content from raw documents. Despite the emergence of numerous DSE systems, their unified evaluation remains inadequate, significantly hindering the field's advancement. This problem is largely attributed to existing benchmark paradigms, which exhibit fragmented and localized characteristics. To address these limitations and offer a thorough evaluation of DSE systems, we introduce a novel benchmark named READoc, which defines DSE as a realistic task of converting unstructured PDFs into semantically rich Markdown. The READoc dataset is derived from 3,576 diverse and real-world documents from arXiv, GitHub, and Zenodo. In addition, we develop a DSE Evaluation S$^3$uite comprising Standardization, Segmentation and Scoring modules, to conduct a unified evaluation of state-of-the-art DSE approaches. By evaluating a range of pipeline tools, expert visual models, and general VLMs, we identify the gap between current work and the unified, realistic DSE objective for the first time. We aspire that READoc will catalyze future research in DSE, fostering more comprehensive and practical solutions.

Boxi Cao, Mengjie Ren, Hongyu Lin et al.

Evaluation is the baton for the development of large language models. Current evaluations typically employ a single-item assessment paradigm for each atomic test objective, which struggles to discern whether a model genuinely possesses the required capabilities or merely memorizes/guesses the answers to specific questions. To this end, we propose a novel evaluation framework referred to as StructEval. Starting from an atomic test objective, StructEval deepens and broadens the evaluation by conducting a structured assessment across multiple cognitive levels and critical concepts, and therefore offers a comprehensive, robust and consistent evaluation for LLMs. Experiments on three widely-used benchmarks demonstrate that StructEval serves as a reliable tool for resisting the risk of data contamination and reducing the interference of potential biases, thereby providing more reliable and consistent conclusions regarding model capabilities. Our framework also sheds light on the design of future principled and trustworthy LLM evaluation protocols.

Xuanang Chen, Ben He, Le Sun et al.

Neural ranking models (NRMs) have undergone significant development and have become integral components of information retrieval (IR) systems. Unfortunately, recent research has unveiled the vulnerability of NRMs to adversarial document manipulations, potentially exploited by malicious search engine optimization practitioners. While progress in adversarial attack strategies aids in identifying the potential weaknesses of NRMs before their deployment, the defensive measures against such attacks, like the detection of adversarial documents, remain inadequately explored. To mitigate this gap, this paper establishes a benchmark dataset to facilitate the investigation of adversarial ranking defense and introduces two types of detection tasks for adversarial documents. A comprehensive investigation of the performance of several detection baselines is conducted, which involve examining the spamicity, perplexity, and linguistic acceptability, and utilizing supervised classifiers. Experimental results demonstrate that a supervised classifier can effectively mitigate known attacks, but it performs poorly against unseen attacks. Furthermore, such classifier should avoid using query text to prevent learning the classification on relevance, as it might lead to the inadvertent discarding of relevant documents.

Tianshu Wang, Hongyu Lin, Cheng Fu et al.

Entity matching (EM) is the most critical step for entity resolution (ER). While current deep learningbased methods achieve very impressive performance on standard EM benchmarks, their realworld application performance is much frustrating. In this paper, we highlight that such the gap between reality and ideality stems from the unreasonable benchmark construction process, which is inconsistent with the nature of entity matching and therefore leads to biased evaluations of current EM approaches. To this end, we build a new EM corpus and re-construct EM benchmarks to challenge critical assumptions implicit in the previous benchmark construction process by step-wisely changing the restricted entities, balanced labels, and single-modal records in previous benchmarks into open entities, imbalanced labels, and multimodal records in an open environment. Experimental results demonstrate that the assumptions made in the previous benchmark construction process are not coincidental with the open environment, which conceal the main challenges of the task and therefore significantly overestimate the current progress of entity matching. The constructed benchmarks and code are publicly released

Xin Zheng, Hongyu Lin, Xianpei Han et al.

Controllable text generation is a fundamental aspect of natural language generation, with numerous methods proposed for different constraint types. However, these approaches often require significant architectural or decoding modifications, making them challenging to apply to additional constraints or resolve different constraint combinations. To address this, our paper introduces Regular Expression Instruction (REI), which utilizes an instruction-based mechanism to fully exploit regular expressions' advantages to uniformly model diverse constraints. Specifically, our REI supports all popular fine-grained controllable generation constraints, i.e., lexical, positional, and length, as well as their complex combinations, via regular expression-style instructions. Our method only requires fine-tuning on medium-scale language models or few-shot, in-context learning on large language models, and requires no further adjustment when applied to various constraint combinations. Experiments demonstrate that our straightforward approach yields high success rates and adaptability to various constraints while maintaining competitiveness in automatic metrics and outperforming most previous baselines.

Ying Zhou, Xuanang Chen, Ben He et al.

Knowledge graph completion (KGC) aims to infer missing knowledge triples based on known facts in a knowledge graph. Current KGC research mostly follows an entity ranking protocol, wherein the effectiveness is measured by the predicted rank of a masked entity in a test triple. The overall performance is then given by a micro(-average) metric over all individual answer entities. Due to the incomplete nature of the large-scale knowledge bases, such an entity ranking setting is likely affected by unlabelled top-ranked positive examples, raising questions on whether the current evaluation protocol is sufficient to guarantee a fair comparison of KGC systems. To this end, this paper presents a systematic study on whether and how the label sparsity affects the current KGC evaluation with the popular micro metrics. Specifically, inspired by the TREC paradigm for large-scale information retrieval (IR) experimentation, we create a relatively "complete" judgment set based on a sample from the popular FB15k-237 dataset following the TREC pooling method. According to our analysis, it comes as a surprise that switching from the original labels to our "complete" labels results in a drastic change of system ranking of a variety of 13 popular KGC models in terms of micro metrics. Further investigation indicates that the IR-like macro(-average) metrics are more stable and discriminative under different settings, meanwhile, less affected by label sparsity. Thus, for KGC evaluation, we recommend conducting TREC-style pooling to balance between human efforts and label completeness, and reporting also the IR-like macro metrics to reflect the ranking nature of the KGC task.

Ruoxi Xu, Hongyu Lin, Meng Liao et al.

Events are considered as the fundamental building blocks of the world. Mining event-centric opinions can benefit decision making, people communication, and social good. Unfortunately, there is little literature addressing event-centric opinion mining, although which significantly diverges from the well-studied entity-centric opinion mining in connotation, structure, and expression. In this paper, we propose and formulate the task of event-centric opinion mining based on event-argument structure and expression categorizing theory. We also benchmark this task by constructing a pioneer corpus and designing a two-step benchmark framework. Experiment results show that event-centric opinion mining is feasible and challenging, and the proposed task, dataset, and baselines are beneficial for future studies.

Shan Wu, Chunlei Xin, Bo Chen et al.

Since the meaning representations are detailed and accurate annotations which express fine-grained sequence-level semtantics, it is usually hard to train discriminative semantic parsers via Maximum Likelihood Estimation (MLE) in an autoregressive fashion. In this paper, we propose a semantic-aware contrastive learning algorithm, which can learn to distinguish fine-grained meaning representations and take the overall sequence-level semantic into consideration. Specifically, a multi-level online sampling algorithm is proposed to sample confusing and diverse instances. Three semantic-aware similarity functions are designed to accurately measure the distance between meaning representations as a whole. And a ranked contrastive loss is proposed to pull the representations of the semantic-identical instances together and push negative instances away. Experiments on two standard datasets show that our approach achieves significant improvements over MLE baselines and gets state-of-the-art performances by simply applying semantic-aware contrastive learning on a vanilla Seq2Seq model.

Zhengzhao Ma, Xueru Wen, Boxi Cao et al.

Reinforcement Learning from Verifiable Rewards (RLVR) significantly enhances large language models (LLMs) reasoning but severely suffers from calibration degeneration, where models become excessively over-confident in incorrect answers. Previous studies devote to directly incorporating calibration objective into existing optimization target. However, our theoretical analysis demonstrates that there exists a fundamental gradient conflict between the optimization for maximizing policy accuracy and minimizing calibration error. Building on this insight, we propose DCPO, a simple yet effective framework that systematically decouples reasoning and calibration objectives. Extensive experiments demonstrate that our DCPO not only preserves accuracy on par with GRPO but also achieves the best calibration performance and substantially mitigates the over-confidence issue. Our study provides valuable insights and practical solution for more reliable LLM deployment.

Qingyu Zhang, Chunlei Xin, Xuanang Chen et al.

Goal-driven persuasive dialogue, exemplified by applications like telemarketing, requires sophisticated multi-turn planning and strict factual faithfulness, which remains a significant challenge for even state-of-the-art Large Language Models (LLMs). A lack of task-specific data often limits previous works, and direct LLM application suffers from strategic brittleness and factual hallucination. In this paper, we first construct and release TeleSalesCorpus, the first real-world-grounded dialogue dataset for this domain. We then propose AI-Salesman, a novel framework featuring a dual-stage architecture. For the training stage, we design a Bayesian-supervised reinforcement learning algorithm that learns robust sales strategies from noisy dialogues. For the inference stage, we introduce the Dynamic Outline-Guided Agent (DOGA), which leverages a pre-built script library to provide dynamic, turn-by-turn strategic guidance. Moreover, we design a comprehensive evaluation framework that combines fine-grained metrics for key sales skills with the LLM-as-a-Judge paradigm. Experimental results demonstrate that our proposed AI-Salesman significantly outperforms baseline models in both automatic metrics and comprehensive human evaluations, showcasing its effectiveness in complex persuasive scenarios.

Shuainan Liu, Xuanang Chen, Ben He et al.

Knowledge editing methods for large language models are commonly evaluated using predefined benchmarks that assess edited facts together with a limited set of related or neighboring knowledge. While effective, such evaluations remain confined to finite, dataset-bounded samples, leaving the broader impact of editing on the model's knowledge system insufficiently understood. To address this gap, we introduce Embedding-Virtualized Knowledge (EVK) that characterizes model knowledge through controlled perturbations in embedding space, enabling the exploration of a substantially broader and virtualized knowledge region beyond explicit data annotations. Based on EVK, we construct an embedding-level evaluation benchmark EVK-Bench that quantifies potential knowledge drift induced by editing, revealing effects that are not captured by conventional sample-based metrics. Furthermore, we propose a plug-and-play EVK-Align module that constrains embedding-level knowledge drift during editing and can be seamlessly integrated into existing editing methods. Experiments demonstrate that our approach enables more comprehensive evaluation while significantly improving knowledge preservation without sacrificing editing accuracy.

Xinyu Lu, Kaiqi Zhang, Jinglin Yang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for enhancing the reasoning capabilities of Large Language Models (LLMs). However, vanilla RLVR suffers from inefficient exploration, particularly when confronting "hard samples" that yield nearzero success rates. In such scenarios, the reliance on sparse outcome rewards typically results in zero-advantage estimates, effectively starving the model of supervision signals despite the high informational value of these instances. To address this, we propose P^2O, a novel framework that synergizes Prompt Optimization with Policy Optimization. P^2O identifies hard samples during training iterations and leverages the GeneticPareto (GEPA) prompt optimization algorithm to evolve prompt templates that guide the model toward discovering successful trajectories. Crucially, unlike traditional prompt engineering methods that rely on input augmentation, P^2O distills the reasoning gains induced by these optimized prompts directly into the model parameters. This mechanism provides denser positive supervision signals for hard samples and accelerates convergence. Extensive experiments demonstrate that P^2O not only achieves superior performance on in-distribution datasets but also exhibits strong generalization, yielding substantial improvements on out-of-distribution benchmarks (+4.7% avg.).

Guozhao Mo, Wenliang Zhong, Jiawei Chen et al.

With the rapid development of Model Context Protocol (MCP), the number of MCP servers has surpassed 10,000. However, existing MCP benchmarks are limited to single-server settings with only a few tools, hindering effective evaluation of agent capabilities in large-scale, real-world scenarios. To address this limitation, we present LiveMCPBench, the first comprehensive benchmark comprising 95 real-world tasks grounded in the MCP ecosystem, designed to evaluate LLM agents at scale across diverse servers. To support a scalable and reproducible evaluation pipeline in large-scale MCP environments, we curate LiveMCPTool, a diverse and readily deployable collection of 70 MCP servers and 527 tools. Furthermore, we introduce LiveMCPEval, an LLM-as-a-Judge framework that enables automated and adaptive evaluation in dynamic, time-varying task environments, achieving 81% agreement with human reviewers. Finally, we propose the MCP Copilot Agent, a multi-step agent that routes tools for dynamic planning and executes tools for API interaction across the entire LiveMCPTool suite. Our evaluation covers 10 leading models, with the best-performing model (Claude-Sonnet-4) reaching a 78.95% success rate. However, we observe large performance variance across models, and several widely-used models perform poorly in LiveMCPBench's complex, tool-rich environments. Overall, LiveMCPBench offers the first unified framework for benchmarking LLM agents in realistic, tool-rich, and dynamic MCP environments, laying a solid foundation for scalable and reproducible research on agent capabilities. Our code and data will be publicly available at https://icip-cas.github.io/LiveMCPBench.

Jialun Cao, Yaojie Lu, Meiziniu Li et al.

The research in AI-based formal mathematical reasoning has shown an unstoppable growth trend. These studies have excelled in mathematical competitions like IMO and have made significant progress. This paper focuses on formal verification, an immediate application scenario of formal reasoning, and breaks it down into sub-tasks. We constructed 18k high-quality instruction-response pairs across five formal specification languages (Coq, Lean4, Dafny, ACSL, and TLA+) by distilling gpt-4o and evaluated against ten open-sourced LLMs, including recent popular DeepSeek-R1. We also fine-tuned several 7~8B small models to achieve comparable performance with Deepseek-R1-671B. Interestingly, we observed that fine-tuning with formal data also enhances mathematics, reasoning, and coding capabilities. Fine-tuned models are released at https: //huggingface.co/fm-universe.

Ying Zhou, Xinyao Wang, Yulei Niu et al.

Recent advancements in large language models (LLMs) have significantly enhanced their knowledge and generative capabilities, leading to a surge of interest in leveraging LLMs for high-quality data synthesis. However, synthetic data generation via prompting LLMs remains challenging due to LLMs' limited understanding of target data distributions and the complexity of prompt engineering, especially for structured formatted data. To address these issues, we introduce DiffLM, a controllable data synthesis framework based on variational autoencoder (VAE), which further (1) leverages diffusion models to reserve more information of original distribution and format structure in the learned latent distribution and (2) decouples the learning of target distribution knowledge from the LLM's generative objectives via a plug-and-play latent feature injection module. As we observed significant discrepancies between the VAE's latent representations and the real data distribution, the latent diffusion module is introduced into our framework to learn a fully expressive latent distribution. Evaluations on seven real-world datasets with structured formatted data (i.e., Tabular, Code, and Tool data) demonstrate that DiffLM generates high-quality data, with performance on downstream tasks surpassing that of real data by 2%-7% in certain cases. Data and code are available at https://github.com/bytedance/DiffLM.

Kevin Robbins, Xiaotong Liu, Yu Wu et al.

Vision-Language Models like CLIP create aligned embedding spaces for text and images, making it possible for anyone to build a visual classifier by simply naming the classes they want to distinguish. However, a model that works well in one domain may fail in another, and non-expert users have no straightforward way to assess whether their chosen VLM will work on their problem. We build on prior work using text-only comparisons to evaluate how well a model works for a given natural language task, and explore approaches that also generate synthetic images relevant to that task to evaluate and refine the prediction of zero-shot accuracy. We show that generated imagery to the baseline text-only scores substantially improves the quality of these predictions. Additionally, it gives a user feedback on the kinds of images that were used to make the assessment. Experiments on standard CLIP benchmark datasets demonstrate that the image-based approach helps users predict, without any labeled examples, whether a VLM will be effective for their application.

Dan Wang, Guozhao Mo, Yafei Shi et al.

Multilingual Retrieval-Augmented Generation (mRAG) leverages cross-lingual evidence to ground Large Language Models (LLMs) in global knowledge. However, we show that current mRAG systems suffer from a language bias during reranking, systematically favoring English and the query's native language. By introducing an estimated oracle evidence analysis, we quantify a substantial performance gap between existing rerankers and the achievable upper bound. Further analysis reveals a critical distributional mismatch: while optimal predictions require evidence scattered across multiple languages, current systems systematically suppress such ``answer-critical'' documents, thereby limiting downstream generation performance. To bridge this gap, we propose \textit{\textbf{L}anguage-\textbf{A}gnostic \textbf{U}tility-driven \textbf{R}eranker \textbf{A}lignment (LAURA)}, which aligns multilingual evidence ranking with downstream generative utility. Experiments across diverse languages and generation models show that LAURA effectively mitigates language bias and consistently improves mRAG performance.

Mengjie Ren, Jie Lou, Boxi Cao et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective paradigm for improving the reasoning capabilities of large language models. However, RLVR training is often hindered by sparse binary rewards and weak credit assignment, resulting in ambiguous optimization signals and underutilization of the useful information embedded in failed trajectories. To address this challenge, we propose Correction-Oriented Policy Optimization (CIPO), a simple and effective extension to RLVR that converts on-policy failed trajectories into correction-oriented supervision, without relying on any external signals. By jointly optimizing correction samples derived from the model's own failed attempts together with the standard RLVR objective, CIPO improves learning effectiveness while explicitly enhancing the model's ability to correct its own errors. Extensive experiments across 11 benchmarks spanning mathematical reasoning and code generation demonstrate that CIPO consistently and significantly outperforms strong baselines in both reasoning and correction performance. Moreover, CIPO yields stronger pass@K gains, indicating that it improves the model's intrinsic reasoning capacity rather than merely redistributing probability mass over existing correct answers.

Qianhao Yuan, Qingyu Zhang, Yanjiang Liu et al.

Multimodal Large Language Models (MLLMs) suffer from high computational costs due to their massive size and the large number of visual tokens. In this paper, we investigate layer-wise redundancy in MLLMs by introducing a novel metric, Layer Contribution (LC), which quantifies the impact of a layer's transformations on visual and text tokens, respectively. The calculation of LC involves measuring the divergence in model output that results from removing the layer's transformations on the specified tokens. Our pilot experiment reveals that many layers of MLLMs exhibit minimal contribution during the processing of visual tokens. Motivated by this observation, we propose ShortV, a training-free method that leverages LC to identify ineffective layers, and freezes visual token updates in these layers. Experiments show that ShortV can freeze visual token in approximately 60\% of the MLLM layers, thereby dramatically reducing computational costs related to updating visual tokens. For example, it achieves a 50\% reduction in FLOPs on LLaVA-NeXT-13B while maintaining superior performance. The code will be publicly available at https://github.com/icip-cas/ShortV

Qiaoyu Tang, Hao Xiang, Le Yu et al.

While recent advances in reasoning models have demonstrated cognitive behaviors through reinforcement learning, existing approaches struggle to invoke deep reasoning capabilities in multi-turn agents with long-horizon interactions. We propose DeepMiner, a novel framework that elicits such abilities by introducing high-difficulty training tasks and dynamic context window. DeepMiner presents a reverse construction method to generate complex but verifiable question-answer pairs from authentic web sources, which ensures the challenge and reliability of training data while injecting cognitive capabilities into multi-turn reasoning scenarios. We further design an elegant yet effective dynamic context management strategy for both training and inference, utilizing sliding window mechanisms while eliminating the dependency on external summarization models, thereby efficiently empowering the model to handle continuously expanding long-horizon contexts. Through reinforcement learning on Qwen3-32B, we develop DeepMiner-32B, which achieves substantial performance improvements across multiple search agent benchmarks. DeepMiner attains 33.5% accuracy on BrowseComp-en, surpassing the previous best open-source agent by almost 20 percentage points, and demonstrates consistent improvements on BrowseComp-zh, XBench-DeepSearch, and GAIA. Notably, our dynamic context management enables sustained interactions of nearly 100 turns within standard 32k context length, effectively addressing the context limitations that constrain existing multi-turn interaction systems.

Duo Wu, Jinghe Wang, Yuan Meng et al.

Utilizing large language models (LLMs) for tool planning has emerged as a promising avenue for developing general AI systems, where LLMs automatically schedule external tools (e.g., vision models) to tackle complex tasks based on task descriptions. To push this paradigm toward practical applications, it is crucial for LLMs to consider tool execution costs (e.g., execution time) for tool planning. Unfortunately, prior studies overlook the tool execution costs, leading to the generation of expensive plans whose costs outweigh their benefits in terms of task performance. To fill this gap, we propose the Cost-Aware Tool Planning with LLMs (CATP-LLM) framework, which for the first time provides a coherent design to empower LLMs for cost-aware tool planning. Specifically, To facilitate efficient concurrent tool execution and cost reduction, we design a tool planning language to enhance the LLM for creating multi-branch non-sequential plans. Moreover, we propose a cost-aware offline reinforcement learning algorithm to fine-tune the LLM to optimize the performance-cost trade-off in tool planning. In the lack of public cost-related datasets, we further present OpenCATP, the first dataset for cost-aware planning, which comprises 11,100 evaluation samples from diverse tasks. Extensive experiments show that CATP-LLM outperforms GPT-4 even when using Llama2-7B as its backbone, with the average improvement of 1.5%-93.9% in terms of plan quality. Codes and dataset are available at: https://github.com/duowuyms/OpenCATP-LLM.

Hao Xiang, Tianyi Tang, Yang Su et al.

Recent advancements in Large Language Models (LLMs) have shown outstanding potential for role-playing applications. Evaluating these capabilities is becoming crucial yet remains challenging. Existing benchmarks mostly adopt a \textbf{character-centric} approach, simplify user-character interactions to isolated Q&A tasks, and fail to reflect real-world applications. To address this limitation, we introduce RMTBench, a comprehensive \textbf{user-centric} bilingual role-playing benchmark featuring 80 diverse characters and over 8,000 dialogue rounds. RMTBench includes custom characters with detailed backgrounds and abstract characters defined by simple traits, enabling evaluation across various user scenarios. Our benchmark constructs dialogues based on explicit user motivations rather than character descriptions, ensuring alignment with practical user applications. Furthermore, we construct an authentic multi-turn dialogue simulation mechanism. With carefully selected evaluation dimensions and LLM-based scoring, this mechanism captures the complex intention of conversations between the user and the character. By shifting focus from character background to user intention fulfillment, RMTBench bridges the gap between academic evaluation and practical deployment requirements, offering a more effective framework for assessing role-playing capabilities in LLMs. All code and datasets will be released soon. We release the datasets at https://huggingface.co/datasets/xiangh/RMTBENCH.

Xueru Wen, Jie Lou, Zichao Li et al.

Reward models (RMs) are crucial for aligning large language models (LLMs) with human preferences. However, most RM research is centered on English and relies heavily on synthetic resources, which leads to limited and less reliable datasets and benchmarks for Chinese. To address this gap, we introduce CheemsBench, a fully human-annotated RM evaluation benchmark within Chinese contexts, and CheemsPreference, a large-scale and diverse preference dataset annotated through human-machine collaboration to support Chinese RM training. We systematically evaluate open-source discriminative and generative RMs on CheemsBench and observe significant limitations in their ability to capture human preferences in Chinese scenarios. Additionally, based on CheemsPreference, we construct an RM that achieves state-of-the-art performance on CheemsBench, demonstrating the necessity of human supervision in RM training. Our findings reveal that scaled AI-generated data struggles to fully capture human preferences, emphasizing the importance of high-quality human supervision in RM development.

Zichao Li, Jie Lou, Fangchen Dong et al.

Reinforcement learning significantly enhances LLM capabilities but suffers from a critical issue: length inflation, where models adopt verbosity or inefficient reasoning to maximize rewards. Prior approaches struggle to address this challenge in a general and lossless manner, primarily because additive penalties introduce a compensatory effect that creates optimization shortcuts, while heuristic gating strategies lack generality beyond binary feedback. To bridge this gap, we present Group Relative Reward Rescaling (GR$^3$), which reframes length control as a multiplicative rescaling paradigm, effectively establishing a generalized, continuous, and reward-dependent gating mechanism. To further ensure lossless optimization, we incorporate group-relative regularization and advantage-aware calibration, which dynamically adapt length budgets to instance difficulty and preserve the advantage signal of high-quality trajectories. Empirically, across both RLHF and RLVR settings, GR$^3$~maintains training dynamics and downstream performance comparable to standard GRPO while significantly mitigating length inflation, outperforming state-of-the-art length-regularized baselines.