Xuezhi Cao

Linyue Pan, Yaoming Zhu, Lin Qiu et al.

Self-improving language agents are typically evaluated in isolation: an agent attempts a task, receives feedback, and iteratively refines its own behavior. Yet agents increasingly operate alongside peers whose strategies and outcomes are publicly visible. This raises an under-studied question: when does shared experience produce improvements that self-improvement alone cannot achieve? We introduce SAGE (Social Agent Group Evolution),an evaluation framework that compares two compute-matched conditions: SocialEvo, where agents from five distinct model families co-evolve with access to all peers' histories; and SelfEvo, where each agent receives the same number of task attempts but sees only its own past, which is conventional in self-improving agent studies. We instantiate SAGE in three arenas: open-ended ML research, long-horizon economic planning, and strategic multiplayer play, evaluated across multiple evolutionary rounds. We find that group history is not a universal amplifier: the strongest agent does not exceed its self-evolution ceiling. However, agents that plateau under self-improvement can achieve significant breakthroughs when peer experience is available. In competitive settings, counterfactual controls reveal that agents improve generally rather than developing opponent-specific strategies. Across different forms of shared history, filtered peer traces and reflective summaries often outperform raw logs, indicating that social gains depend on abstraction rather than exposure volume. These findings reveal that peer-history gains are agent-specific, arena-dependent, and contingent on the capacity to abstract transferable knowledge from public traces.

Jiawei Chen, Junkang Wu, Jiancan Wu et al.

Recent years have witnessed the great successes of embedding-based methods in recommender systems. Despite their decent performance, we argue one potential limitation of these methods -- the embedding magnitude has not been explicitly modulated, which may aggravate popularity bias and training instability, hindering the model from making a good recommendation. It motivates us to leverage the embedding normalization in recommendation. By normalizing user/item embeddings to a specific value, we empirically observe impressive performance gains (9\% on average) on four real-world datasets. Although encouraging, we also reveal a serious limitation when applying normalization in recommendation -- the performance is highly sensitive to the choice of the temperature $τ$ which controls the scale of the normalized embeddings. To fully foster the merits of the normalization while circumvent its limitation, this work studied on how to adaptively set the proper $τ$. Towards this end, we first make a comprehensive analyses of $τ$ to fully understand its role on recommendation. We then accordingly develop an adaptive fine-grained strategy Adap-$τ$ for the temperature with satisfying four desirable properties including adaptivity, personalized, efficiency and model-agnostic. Extensive experiments have been conducted to validate the effectiveness of the proposal. The code is available at \url{https://github.com/junkangwu/Adap_tau}.

Peng Ding, Jun Kuang, Dan Ma et al.

Large Language Models (LLMs), such as ChatGPT and GPT-4, are designed to provide useful and safe responses. However, adversarial prompts known as 'jailbreaks' can circumvent safeguards, leading LLMs to generate potentially harmful content. Exploring jailbreak prompts can help to better reveal the weaknesses of LLMs and further steer us to secure them. Unfortunately, existing jailbreak methods either suffer from intricate manual design or require optimization on other white-box models, which compromises either generalization or efficiency. In this paper, we generalize jailbreak prompt attacks into two aspects: (1) Prompt Rewriting and (2) Scenario Nesting. Based on this, we propose ReNeLLM, an automatic framework that leverages LLMs themselves to generate effective jailbreak prompts. Extensive experiments demonstrate that ReNeLLM significantly improves the attack success rate while greatly reducing the time cost compared to existing baselines. Our study also reveals the inadequacy of current defense methods in safeguarding LLMs. Finally, we analyze the failure of LLMs defense from the perspective of prompt execution priority, and propose corresponding defense strategies. We hope that our research can catalyze both the academic community and LLMs developers towards the provision of safer and more regulated LLMs. The code is available at https://github.com/NJUNLP/ReNeLLM.

Deli Huang, Cunguang Wang, Hongyin Tang et al.

Long-context language models now advertise context windows up to millions of tokens, yet evaluations typically report a single length or a narrow task family, masking two failure modes: performance can collapse as length grows, and strong retrieval need not transfer to downstream use. We present ATLAS, a benchmarking framework that redefines long-context evaluation as length-dependent capability profiling. ATLAS contributes three methodological principles:(i) a layered taxonomy separating foundational operations from application workloads so failures can be attributed, (ii) length-aware AUC scoring that integrates score-length curves over a fixed 8K-1M grid, replacing single-point metrics with full degradation profiles, and (iii) ATLAScore, a harmonic-mean aggregate over taxonomy categories that penalizes imbalanced profiles, with end-to-end uncertainty propagation from subset scores through the nonlinear final aggregate. We instantiate the framework across eight capability dimensions with nine auditable components and 6,438 instances, and evaluate 26 models. Gemini-3.1-Pro-Preview leads at 128K, Claude-Opus-4.6 leads at 1M. Rankings reshuffle substantially between ATLASscore@8K-128K and ATLASscore@8K-1M: 7 models move by at least two ranks, and the two taxonomy layers share only 61% of cross-model variance, with individual rank gaps up to 12 positions. These results support reporting long-context quality by capability and length, not by a single headline score.

Renyu Zhu, Chengcheng Han, Yong Qian et al. · stanford

We study the problem of multimodal fusion in this paper. Recent exchanging-based methods have been proposed for vision-vision fusion, which aim to exchange embeddings learned from one modality to the other. However, most of them project inputs of multimodalities into different low-dimensional spaces and cannot be applied to the sequential input data. To solve these issues, in this paper, we propose a novel exchanging-based multimodal fusion model MuSE for text-vision fusion based on Transformer. We first use two encoders to separately map multimodal inputs into different low-dimensional spaces. Then we employ two decoders to regularize the embeddings and pull them into the same space. The two decoders capture the correlations between texts and images with the image captioning task and the text-to-image generation task, respectively. Further, based on the regularized embeddings, we present CrossTransformer, which uses two Transformer encoders with shared parameters as the backbone model to exchange knowledge between multimodalities. Specifically, CrossTransformer first learns the global contextual information of the inputs in the shallow layers. After that, it performs inter-modal exchange by selecting a proportion of tokens in one modality and replacing their embeddings with the average of embeddings in the other modality. We conduct extensive experiments to evaluate the performance of MuSE on the Multimodal Named Entity Recognition task and the Multimodal Sentiment Analysis task. Our results show the superiority of MuSE against other competitors. Our code and data are provided at https://github.com/RecklessRonan/MuSE.

Peng Ding, Jingyu Wu, Jun Kuang et al.

Multi-modal Large Language Models (MLLMs) have demonstrated remarkable performance on various visual-language understanding and generation tasks. However, MLLMs occasionally generate content inconsistent with the given images, which is known as "hallucination". Prior works primarily center on evaluating hallucination using standard, unperturbed benchmarks, which overlook the prevalent occurrence of perturbed inputs in real-world scenarios-such as image cropping or blurring-that are critical for a comprehensive assessment of MLLMs' hallucination. In this paper, to bridge this gap, we propose Hallu-PI, the first benchmark designed to evaluate Hallucination in MLLMs within Perturbed Inputs. Specifically, Hallu-PI consists of seven perturbed scenarios, containing 1,260 perturbed images from 11 object types. Each image is accompanied by detailed annotations, which include fine-grained hallucination types, such as existence, attribute, and relation. We equip these annotations with a rich set of questions, making Hallu-PI suitable for both discriminative and generative tasks. Extensive experiments on 12 mainstream MLLMs, such as GPT-4V and Gemini-Pro Vision, demonstrate that these models exhibit significant hallucinations on Hallu-PI, which is not observed in unperturbed scenarios. Furthermore, our research reveals a severe bias in MLLMs' ability to handle different types of hallucinations. We also design two baselines specifically for perturbed scenarios, namely Perturbed-Reminder and Perturbed-ICL. We hope that our study will bring researchers' attention to the limitations of MLLMs when dealing with perturbed inputs, and spur further investigations to address this issue. Our code and datasets are publicly available at https://github.com/NJUNLP/Hallu-PI.

Kaining Ying, Hengrui Hu, Siyu Ren et al.

Interactive world models are advancing rapidly, yet existing benchmarks cover only part of the required competencies, leaving no unified standard for systematic evaluation. To fill this gap, we introduce WBench, a comprehensive multi-turn benchmark for interactive world model evaluation along five dimensions, namely video quality, setting adherence, interaction adherence, consistency, and physics compliance. WBench contains 289 test cases and 1,058 interaction turns, where each case specifies a world setting and a multi-turn interaction sequence, covering diverse scenes, styles, subjects, and both first- and third-person perspectives, together with four interaction types, including navigation, subject action, event editing, and perspective switching. For navigation, WBench unifies text, 6-DoF pose, and discrete-action control, enabling evaluation of models with different native input interfaces. Evaluation uses 22 automatic sub-metrics that combine specialist vision models with large multimodal models, and all metrics are validated against human judgments. Across 20 state-of-the-art models, we find that no single model performs strongly across all dimensions. We provide detailed diagnostic insights into the characteristic strengths, weaknesses, and open challenges of each model. Code and data are available at https://github.com/meituan-longcat/WBench.

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We introduce LongCat-Flash-Thinking-2601, a 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model with superior agentic reasoning capability. LongCat-Flash-Thinking-2601 achieves state-of-the-art performance among open-source models on a wide range of agentic benchmarks, including agentic search, agentic tool use, and tool-integrated reasoning. Beyond benchmark performance, the model demonstrates strong generalization to complex tool interactions and robust behavior under noisy real-world environments. Its advanced capability stems from a unified training framework that combines domain-parallel expert training with subsequent fusion, together with an end-to-end co-design of data construction, environments, algorithms, and infrastructure spanning from pre-training to post-training. In particular, the model's strong generalization capability in complex tool-use are driven by our in-depth exploration of environment scaling and principled task construction. To optimize long-tailed, skewed generation and multi-turn agentic interactions, and to enable stable training across over 10,000 environments spanning more than 20 domains, we systematically extend our asynchronous reinforcement learning framework, DORA, for stable and efficient large-scale multi-environment training. Furthermore, recognizing that real-world tasks are inherently noisy, we conduct a systematic analysis and decomposition of real-world noise patterns, and design targeted training procedures to explicitly incorporate such imperfections into the training process, resulting in improved robustness for real-world applications. To further enhance performance on complex reasoning tasks, we introduce a Heavy Thinking mode that enables effective test-time scaling by jointly expanding reasoning depth and width through intensive parallel thinking.

Meituan LongCat Team, Bin Xiao, Chao Wang et al.

The prevailing Next-Token Prediction (NTP) paradigm has driven the success of large language models through discrete autoregressive modeling. However, contemporary multimodal systems remain language-centric, often treating non-linguistic modalities as external attachments, leading to fragmented architectures and suboptimal integration. To transcend this limitation, we introduce Discrete Native Autoregressive (DiNA), a unified framework that represents multimodal information within a shared discrete space, enabling a consistent and principled autoregressive modeling across modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which performs tokenization and de-tokenization at arbitrary resolutions, transforming continuous visual signals into hierarchical discrete tokens. Building on this foundation, we develop LongCat-Next, a native multimodal model that processes text, vision, and audio under a single autoregressive objective with minimal modality-specific design. As an industrial-strength foundation model, it excels at seeing, painting, and talking within a single framework, achieving strong performance across a wide range of multimodal benchmarks. In particular, LongCat-Next addresses the long-standing performance ceiling of discrete vision modeling on understanding tasks and provides a unified approach to effectively reconcile the conflict between understanding and generation. As an attempt toward native multimodality, we open-source the LongCat-Next and its tokenizers, hoping to foster further research and development in the community. GitHub: https://github.com/meituan-longcat/LongCat-Next

Chengcheng Han, Renyu Zhu, Jun Kuang et al.

Meta-learning methods have been widely used in few-shot named entity recognition (NER), especially prototype-based methods. However, the Other(O) class is difficult to be represented by a prototype vector because there are generally a large number of samples in the class that have miscellaneous semantics. To solve the problem, we propose MeTNet, which generates prototype vectors for entity types only but not O-class. We design an improved triplet network to map samples and prototype vectors into a low-dimensional space that is easier to be classified and propose an adaptive margin for each entity type. The margin plays as a radius and controls a region with adaptive size in the low-dimensional space. Based on the regions, we propose a new inference procedure to predict the label of a query instance. We conduct extensive experiments in both in-domain and cross-domain settings to show the superiority of MeTNet over other state-of-the-art methods. In particular, we release a Chinese few-shot NER dataset FEW-COMM extracted from a well-known e-commerce platform. To the best of our knowledge, this is the first Chinese few-shot NER dataset. All the datasets and codes are provided at https://github.com/hccngu/MeTNet.

Wentao Shi, Junkang Wu, Xuezhi Cao et al.

Learning hyperbolic embeddings for knowledge graph (KG) has gained increasing attention due to its superiority in capturing hierarchies. However, some important operations in hyperbolic space still lack good definitions, making existing methods unable to fully leverage the merits of hyperbolic space. Specifically, they suffer from two main limitations: 1) existing Graph Convolutional Network (GCN) methods in hyperbolic space rely on tangent space approximation, which would incur approximation error in representation learning, and 2) due to the lack of inner product operation definition in hyperbolic space, existing methods can only measure the plausibility of facts (links) with hyperbolic distance, which is difficult to capture complex data patterns. In this work, we contribute: 1) a Full Poincaré Multi-relational GCN that achieves graph information propagation in hyperbolic space without requiring any approximation, and 2) a hyperbolic generalization of Euclidean inner product that is beneficial to capture both hierarchical and complex patterns. On this basis, we further develop a \textbf{F}ully and \textbf{F}lexible \textbf{H}yperbolic \textbf{R}epresentation framework (\textbf{FFHR}) that is able to transfer recent Euclidean-based advances to hyperbolic space. We demonstrate it by instantiating FFHR with four representative KGC methods. Extensive experiments on benchmark datasets validate the superiority of our FFHRs over their Euclidean counterparts as well as state-of-the-art hyperbolic embedding methods.

Jianing Wang, Jianfei Zhang, Qi Guo et al.

We introduce LongCat-Flash-Prover, a flagship 560-billion-parameter open-source Mixture-of- Experts (MoE) model that advances Native Formal Reasoning in Lean4 through agentic tool-integrated reasoning (TIR). We decompose the native formal reasoning task into three independent formal capabilities, i.e., auto-formalization, sketching, and proving. To facilitate these capabilities, we propose a Hybrid-Experts Iteration Framework to expand high-quality task trajectories, including generating a formal statement based on a given informal problem, producing a whole-proof directly from the statement, or a lemma-style sketch. During agentic RL, we present a Hierarchical Importance Sampling Policy Optimization (HisPO) algorithm, which aims to stabilize the MoE model training on such long-horizon tasks. It employs a gradient masking strategy that accounts for the policy staleness and the inherent train-inference engine discrepancies at both sequence and token levels. Additionally, we also incorporate theorem consistency and legality detection mechanisms to eliminate reward hacking issues. Extensive evaluations show that our LongCat-Flash-Prover sets a new state-of-the-art for open-weights models in both auto-formalization and theorem proving. Demonstrating remarkable sample efficiency, it achieves a 97.1% pass rate on MiniF2F-Test using only 72 inference budget per problem. On more challenging benchmarks, it solves 70.8% of ProverBench and 41.5% of PutnamBench with no more than 220 attempts per problem, significantly outperforming existing open-weights baselines.

Wenxuan Jiang, Yuxin Zuo, Zijian Zhang et al.

In-Context Reinforcement Learning (ICRL) enables Large Language Models (LLMs) to learn online from external rewards directly within the context window. However, a central challenge in ICRL is reward estimation, as models typically lack access to ground-truths during inference. To address this limitation, we propose Test-Time Rethinking for In-Context Reinforcement Learning (TR-ICRL), a novel ICRL framework designed for both reasoning and knowledge-intensive tasks. TR-ICRL operates by first retrieving the most relevant instances from an unlabeled evaluation set for a given query. During each ICRL iteration, LLM generates a set of candidate answers for every retrieved instance. Next, a pseudo-label is derived from this set through majority voting. This label then serves as a proxy to give reward messages and generate formative feedbacks, guiding LLM through iterative refinement. In the end, this synthesized contextual information is integrated with the original query to form a comprehensive prompt, with the answer determining through a final round of majority voting. TR-ICRL is evaluated on mainstream reasoning and knowledge-intensive tasks, where it demonstrates significant performance gains. Remarkably, TR-ICRL improves Qwen2.5-7B by 21.23% on average on MedQA and even 137.59% on AIME2024. Extensive ablation studies and analyses further validate the effectiveness and robustness of our approach. Our code is available at https://github.com/pangpang-xuan/TR_ICRL.

Yingjie Zhou, Xilei Zhu, Siyu Ren et al.

Speech-driven Talking Human (TH) generation, commonly known as "Talker," currently faces limitations in multi-subject driving capabilities. Extending this paradigm to "Multi-Talker," capable of animating multiple subjects simultaneously, introduces richer interactivity and stronger immersion in audiovisual communication. However, current Multi-Talkers still exhibit noticeable quality degradation caused by technical limitations, resulting in suboptimal user experiences. To address this challenge, we construct THQA-MT, the first large-scale Multi-Talker-generated Talking Human Quality Assessment dataset, consisting of 5,492 Multi-Talker-generated THs (MTHs) from 15 representative Multi-Talkers using 400 real portraits collected online. Through subjective experiments, we analyze perceptual discrepancies among different Multi-Talkers and identify 12 common types of distortion. Furthermore, we introduce EvalTalker, a novel TH quality assessment framework. This framework possesses the ability to perceive global quality, human characteristics, and identity consistency, while integrating Qwen-Sync to perceive multimodal synchrony. Experimental results demonstrate that EvalTalker achieves superior correlation with subjective scores, providing a robust foundation for future research on high-quality Multi-Talker generation and evaluation.

Cheng Jiayang, Dongyu Ru, Lin Qiu et al.

Long-horizon interactions between users and LLM-based assistants necessitate effective memory management, yet current approaches face challenges in training and evaluation of memory. Existing memory benchmarks rely on static, off-policy data as context, limiting evaluation reliability and scalability. To address these gaps, we introduce AMemGym, an interactive environment enabling on-policy evaluation and optimization for memory-driven personalization. AMemGym employs structured data sampling to predefine user profiles, state-dependent questions, and state evolution trajectories, enabling cost-effective generation of high-quality, evaluation-aligned interactions. LLM-simulated users expose latent states through role-play while maintaining structured state consistency. Comprehensive metrics based on structured data guide both assessment and optimization of assistants. Extensive experiments reveal performance gaps in existing memory systems (e.g., RAG, long-context LLMs, and agentic memory) and corresponding reasons. AMemGym not only enables effective selection among competing approaches but also can potentially drive the self-evolution of memory management strategies. By bridging structured state evolution with free-form interactions, our framework provides a scalable, diagnostically rich environment for advancing memory capabilities in conversational agents.

Meituan LongCat Team, Bairui Wang, Bayan et al.

We introduce LongCat-Flash-Omni, a state-of-the-art open-source omni-modal model with 560 billion parameters, excelling at real-time audio-visual interaction. By adopting a curriculum-inspired progressive training strategy that transitions from simpler to increasingly complex modality sequence modeling tasks, LongCat-Flash-Omni attains comprehensive multimodal capabilities while maintaining strong unimodal capability. Building upon LongCat-Flash, which adopts a high-performance Shortcut-connected Mixture-of-Experts (MoE) architecture with zero-computation experts, LongCat-Flash-Omni integrates efficient multimodal perception and speech reconstruction modules. Despite its immense size of 560B parameters (with 27B activated), LongCat-Flash-Omni achieves low-latency real-time audio-visual interaction. For training infrastructure, we developed a modality-decoupled parallelism scheme specifically designed to manage the data and model heterogeneity inherent in large-scale multimodal training. This innovative approach demonstrates exceptional efficiency by sustaining over 90% of the throughput achieved by text-only training. Extensive evaluations show that LongCat-Flash-Omni achieves state-of-the-art performance on omni-modal benchmarks among open-source models. Furthermore, it delivers highly competitive results across a wide range of modality-specific tasks, including text, image, and video understanding, as well as audio understanding and generation. We provide a comprehensive overview of the model architecture design, training procedures, and data strategies, and open-source the model to foster future research and development in the community.

Wei Liu, Yang Bai, Chengcheng Han et al.

Direct Preference Optimization (DPO) is widely utilized in the Reinforcement Learning from Human Feedback (RLHF) phase to align Large Language Models (LLMs) with human preferences, thereby enhancing both their harmlessness and efficacy. However, it has been observed that DPO tends to over-optimize for verbosity, which can detrimentally affect both performance and user experience. In this paper, we conduct an in-depth theoretical analysis of DPO's optimization objective and reveal a strong correlation between its implicit reward and data length. This correlation misguides the optimization direction, resulting in length sensitivity during the DPO training and leading to verbosity. To address this issue, we propose a length-desensitization improvement method for DPO, termed LD-DPO. The proposed method aims to desensitize DPO to data length by decoupling explicit length preference, which is relatively insignificant, from the other implicit preferences, thereby enabling more effective learning of the intrinsic preferences. We utilized two settings (Base and Instruct) of Llama2-13B, Llama3-8B, and Qwen2-7B for experimental validation on various benchmarks including MT-Bench and AlpacaEval 2. The experimental results indicate that LD-DPO consistently outperforms DPO and other baseline methods, achieving more concise responses with a 10-40% reduction in length compared to DPO. We conducted in-depth experimental analyses to demonstrate that LD-DPO can indeed achieve length desensitization and align the model more closely with human-like preferences.

Junlin Liu, Shengnan An, Shuang Zhou et al.

Contemporary large language models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in specialized domains like mathematics and physics. However, their ability to generalize these reasoning skills to more general and broader contexts--often termed general reasoning--remains under-explored. Unlike domain-specific reasoning, general reasoning relies less on expert knowledge but still presents formidable reasoning challenges, such as complex constraints, nested logical branches, and semantic interference. To address this gap, we introduce General365, a benchmark specifically designed to assess general reasoning in LLMs. By restricting background knowledge to a K-12 level, General365 explicitly decouples reasoning from specialized expertise. The benchmark comprises 365 seed problems and 1,095 variant problems across eight categories, ensuring both high difficulty and diversity. Evaluations across 26 leading LLMs reveal that even the top-performing model achieves only 62.8% accuracy, in stark contrast to the near-perfect performances of LLMs in math and physics benchmarks. These results suggest that the reasoning abilities of current LLMs are heavily domain-dependent, leaving significant room for improvement in broader applications. We envision General365 as a catalyst for advancing LLM reasoning beyond domain-specific tasks toward robust, general-purpose real-world scenarios. Code, Dataset, and Leaderboard: https://general365.github.io

Zicheng Zhang, Tengchuan Kou, Shushi Wang et al.

Evaluating text-to-vision content hinges on two crucial aspects: visual quality and alignment. While significant progress has been made in developing objective models to assess these dimensions, the performance of such models heavily relies on the scale and quality of human annotations. According to Scaling Law, increasing the number of human-labeled instances follows a predictable pattern that enhances the performance of evaluation models. Therefore, we introduce a comprehensive dataset designed to Evaluate Visual quality and Alignment Level for text-to-vision content (Q-EVAL-100K), featuring the largest collection of human-labeled Mean Opinion Scores (MOS) for the mentioned two aspects. The Q-EVAL-100K dataset encompasses both text-to-image and text-to-video models, with 960K human annotations specifically focused on visual quality and alignment for 100K instances (60K images and 40K videos). Leveraging this dataset with context prompt, we propose Q-Eval-Score, a unified model capable of evaluating both visual quality and alignment with special improvements for handling long-text prompt alignment. Experimental results indicate that the proposed Q-Eval-Score achieves superior performance on both visual quality and alignment, with strong generalization capabilities across other benchmarks. These findings highlight the significant value of the Q-EVAL-100K dataset. Data and codes will be available at https://github.com/zzc-1998/Q-Eval.

Lingyue Fu, Xin Ding, Yaoming Zhu et al.

Large Language Model (LLM) agents have evolved from basic text generation to autonomously completing complex tasks through interaction with external tools. However, current benchmarks mainly assess end-to-end performance in fixed scenarios, restricting evaluation to specific skills and suffering from score saturation and growing dependence on expert annotation as agent capabilities improve. In this work, we emphasize the importance of learning ability, including both self-improvement and peer-learning, as a core driver for agent evolution toward human-level intelligence. We propose an iterative, competitive peer-learning framework, which allows agents to refine and optimize their strategies through repeated interactions and feedback, thereby systematically evaluating their learning capabilities. To address the score saturation issue in current benchmarks, we introduce CATArena, a tournament-style evaluation platform featuring four diverse board and card games with open-ended scoring. By providing tasks without explicit upper score limits, CATArena enables continuous and dynamic evaluation of rapidly advancing agent capabilities. Experimental results and analyses involving both minimal and commercial code agents demonstrate that CATArena provides reliable, stable, and scalable benchmarking for core agent abilities, particularly learning ability and strategy coding.

Meituan LongCat Team, Bayan, Bei Li et al.

We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat

Shao Zhang, Xihuai Wang, Wenhao Zhang et al.

Agents built on large language models (LLMs) have excelled in turn-by-turn human-AI collaboration but struggle with simultaneous tasks requiring real-time interaction. Latency issues and the challenge of inferring variable human strategies hinder their ability to make autonomous decisions without explicit instructions. Through experiments with current independent System 1 and System 2 methods, we validate the necessity of using Dual Process Theory (DPT) in real-time tasks. We propose DPT-Agent, a novel language agent framework that integrates System 1 and System 2 for efficient real-time simultaneous human-AI collaboration. DPT-Agent's System 1 uses a Finite-state Machine (FSM) and code-as-policy for fast, intuitive, and controllable decision-making. DPT-Agent's System 2 integrates Theory of Mind (ToM) and asynchronous reflection to infer human intentions and perform reasoning-based autonomous decisions. We demonstrate the effectiveness of DPT-Agent through further experiments with rule-based agents and human collaborators, showing significant improvements over mainstream LLM-based frameworks. DPT-Agent can effectively help LLMs convert correct slow thinking and reasoning into executable actions, thereby improving performance. To the best of our knowledge, DPT-Agent is the first language agent framework that achieves successful real-time simultaneous human-AI collaboration autonomously. Code of DPT-Agent can be found in https://github.com/sjtu-marl/DPT-Agent.

Shengnan An, Xunliang Cai, Xuezhi Cao et al.

We present AMO-Bench, an Advanced Mathematical reasoning benchmark with Olympiad level or even higher difficulty, comprising 50 human-crafted problems. Existing benchmarks have widely leveraged high school math competitions for evaluating mathematical reasoning capabilities of large language models (LLMs). However, many existing math competitions are becoming less effective for assessing top-tier LLMs due to performance saturation (e.g., AIME24/25). To address this, AMO-Bench introduces more rigorous challenges by ensuring all 50 problems are (1) cross-validated by experts to meet at least the International Mathematical Olympiad (IMO) difficulty standards, and (2) entirely original problems to prevent potential performance leakages from data memorization. Moreover, each problem in AMO-Bench requires only a final answer rather than a proof, enabling automatic and robust grading for evaluation. Experimental results across 26 LLMs on AMO-Bench show that even the best-performing model achieves only 52.4% accuracy on AMO-Bench, with most LLMs scoring below 40%. Beyond these poor performances, our further analysis reveals a promising scaling trend with increasing test-time compute on AMO-Bench. These results highlight the significant room for improving the mathematical reasoning in current LLMs. We release AMO-Bench to facilitate further research into advancing the reasoning abilities of language models. https://amo-bench.github.io/

Dujun Nie, Fengjiao Chen, Qi Lv et al.

While the shortage of explicit action data limits Vision-Language-Action (VLA) models, human action videos offer a scalable yet unlabeled data source. A critical challenge in utilizing large-scale human video datasets lies in transforming visual signals into ontology-independent representations, known as latent actions. However, the capacity of latent action representation to derive robust control from visual observations has yet to be rigorously evaluated. We introduce the Latent Action Representation Yielding (LARY) Benchmark, a unified framework for evaluating latent action representations on both high-level semantic actions (what to do) and low-level robotic control (how to do). The comprehensively curated dataset encompasses over one million videos (1,000 hours) spanning 151 action categories, alongside 620K image pairs and 595K motion trajectories across diverse embodiments and environments. Our experiments reveal two crucial insights: (i) General visual foundation models, trained without any action supervision, consistently outperform specialized embodied latent action models. (ii) Latent-based visual space is fundamentally better aligned to physical action space than pixel-based space. These results suggest that general visual representations inherently encode action-relevant knowledge for physical control, and that semantic-level abstraction serves as a fundamentally more effective pathway from vision to action than pixel-level reconstruction.

Hao Guan, Lingyue Fu, Shao Zhang et al.

As autonomous code agents move toward end-to-end software development, evaluating their practical autonomy becomes critical. Current benchmarks hide friction by testing agents in pre-configured environments, and their static evaluation pipelines frequently fail when parsing fully autonomous trajectories. We address these limitations with SWE-Cycle, a benchmark of 489 rigorously filtered instances. SWE-Cycle evaluates agents across three isolated tasks, including environment reconstruction, code implementation, and verification test generation, as well as an end-to-end FullCycle task that integrates all three. The FullCycle task requires agents to work autonomously in a bare repository without human scaffolding. To reliably assess these complex execution paths, we developed SWE-Judge. By combining static code review with dynamic testing, this execution-capable evaluation agent accurately verifies functional correctness and eliminates the systematic measurement errors of traditional static parsers. We evaluate code agents powered by six state-of-the-art LLMs across these four tasks. The results reveal a sharp drop in solve rates when transitioning from isolated tasks to FullCycle execution, exposing critical bottlenecks in handling cross-phase dependencies and maintaining code quality. Together, SWE-Cycle and SWE-Judge provide a comprehensive framework for accurately measuring the end-to-end capabilities of autonomous software agents.

Lingyue Fu, Hao Guan, Bolun Zhang et al.

As Large Language Models (LLMs) demonstrate increasingly sophisticated code processing capabilities, evaluating their performance on engineering-level code remains challenging. Existing repository-level benchmarks primarily focus on single scenarios, such as code generation or bug fixing, without adequately capturing the diversity and complexity of real-world software or project engineering workflows. Furthermore, these benchmarks suffer from limited controllability in question positioning and reliability issues in their generated test cases. To address these limitations, we present CorePipe, a fully automated pipeline that converts repositories into comprehensive test cases, and introduce CoreCodeBench, a configurable multi-scenario repository-level benchmark. To simulate real engineering scenarios, CorePipe generates three types of atomic questions (Development, BugFix, and Test-Driven Development) specifically targeting core code segments. These atomic questions are further combined into three types of composite questions, with difficulty levels flexibly adjusted through hyperparameter tuning. CoreCodeBench provides a comprehensive and extensive repository-level benchmark to investigate the applicability of LLMs in real-world engineering projects. Experiments with 16 LLMs across diverse scenarios reveal varying capabilities and offer multi-dimensional insights into LLM performance in engineering contexts. The code for CorePipe is available at https://github.com/AGI-Eval-Official/CoreCodeBench, and the data for CoreCodeBench can be accessed at https://huggingface.co/collections/tubehhh/corecodebench-68256d2faabf4b1610a08caa.

Yaoming Zhu, Junxin Wang, Yiyang Li et al.

As models become increasingly sophisticated, conventional algorithm benchmarks are increasingly saturated, underscoring the need for more challenging benchmarks to guide future improvements in algorithmic reasoning. This paper introduces OIBench, a high-quality, private, and challenging olympiad-level informatics dataset comprising 250 carefully curated original problems. We detail the construction methodology of the benchmark, ensuring a comprehensive assessment across various programming paradigms and complexities, and we demonstrate its contamination-resistant properties via experiments. We propose Time/Space Completion Curves for finer-grained efficiency analysis and enable direct human-model comparisons through high-level participant evaluations. Our experiments reveal that while open-source models lag behind closed-source counterparts, current SOTA models already outperform most human participants in both correctness and efficiency, while still being suboptimal compared to the canonical solutions. By releasing OIBench as a fully open-source resource (https://huggingface.co/datasets/AGI-Eval/OIBench), we hope this benchmark will contribute to advancing code reasoning capabilities for future LLMs.

Peng Ding, Jun Kuang, Zongyu Wang et al.

Large Language Models (LLMs) have shown impressive capabilities across various tasks but remain vulnerable to meticulously crafted jailbreak attacks. In this paper, we identify a critical safety gap: while LLMs are adept at detecting jailbreak prompts, they often produce unsafe responses when directly processing these inputs. Inspired by this insight, we propose SAGE (Self-Aware Guard Enhancement), a training-free defense strategy designed to align LLMs' strong safety discrimination performance with their relatively weaker safety generation ability. SAGE consists of two core components: a Discriminative Analysis Module and a Discriminative Response Module, enhancing resilience against sophisticated jailbreak attempts through flexible safety discrimination instructions. Extensive experiments demonstrate SAGE's effectiveness and robustness across various open-source and closed-source LLMs of different sizes and architectures, achieving an average 99% defense success rate against numerous complex and covert jailbreak methods while maintaining helpfulness on general benchmarks. We further conduct mechanistic interpretability analysis through hidden states and attention distributions, revealing the underlying mechanisms of this detection-generation discrepancy. Our work thus contributes to developing future LLMs with coherent safety awareness and generation behavior. Our code and datasets are publicly available at https://github.com/NJUNLP/SAGE.

Jiaming wang, Yunke Zhao, Peng Ding et al.

The capability to precisely adhere to instructions is a cornerstone for Large Language Models (LLMs) to function as dependable agents in real-world scenarios. However, confronted with complex prompts, LLMs frequently encounter difficulties in fulfilling all specified requirements within a single response. Drawing inspiration from recent advancements in Chain-of-Thought (CoT) prompting and self-correction methodologies, we introduce Meeseeks (The name is inspired by Mr. Meeseeks from "Rick and Morty," a character renowned for efficiently accomplishing assigned tasks. See: https://en.wikipedia.org/wiki/Mr._Meeseeks), a fully automated iterative instruction-following benchmark equipped with an integrated feedback mechanism. Meeseeks identifies erroneous components in model responses and provides corresponding feedback accurately, thereby iteratively guiding the model toward self-correction. The dataset contains over 700 curated instances annotated by 32 distinct capability tags in Chinese and English. Extensive experimental results reveal that different state-of-the-art commercial and open-source LLMs exhibit vastly disparate performance, and even after 20 turns of iterative feedback-driven self-correction, nearly all models demonstrate suboptimal performance. We conducted comprehensive analysis from both macro and instance levels, uncovering numerous common issues prevalent in current state-of-the-art models, as well as several counterintuitive phenomena. We've open-sourced our work on https://github.com/ADoublLEN/Meeseeks.

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We present LongCat-Flash-Thinking, an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model. Its advanced capabilities are cultivated through a meticulously crafted training process, beginning with long Chain-of-Thought (CoT) data cold-start and culminating in large-scale Reinforcement Learning (RL). We first employ a well-designed cold-start training strategy, which significantly enhances the reasoning potential and equips the model with specialized skills in both formal and agentic reasoning. Then, a core innovation is our domain-parallel training scheme, which decouples optimization across distinct domains (e.g., STEM, Code, Agentic) and subsequently fuses the resulting expert models into a single, nearly Pareto-optimal model. This entire process is powered by our Dynamic ORchestration for Asynchronous rollout (DORA) system, a large-scale RL framework that delivers a greater than threefold training speedup over synchronous methods on tens of thousands of accelerators. As a result, LongCat-Flash-Thinking achieves state-of-the-art performance among open-source models on a suite of complex reasoning tasks. The model exhibits exceptional efficiency in agentic reasoning, reducing average token consumption by 64.5% (from 19, 653 to 6, 965) on AIME-25, without degrading task accuracy. We release LongCat-Flash-Thinking to promote further advances in reasoning systems and agentic AI research.

Fengjiao Chen, Minhao Jing, Weitao Lu et al.

Vision-language large models are moving toward the unification of visual understanding and visual generation tasks. However, whether generation can enhance understanding is still under-explored on large data scale. In this work, we analysis the unified structure with a concise model, UniHetero, under large-scale pretraining (>200M samples). Our key observations are: (1) Generation can improve understanding, but Only if you generate Semantics, Not Pixels. A common assumption in unified vision-language models is that adding generation will naturally strengthen understanding. However, this is not always true at scale. At 200M+ pretraining samples, generation helps understanding only when it operates at the semantic level, i.e. when the model learns to autoregress high-level visual representations inside the LLM. Once pixel-level objectives (e.g., diffusion losses) directly interfere with the LLM, understanding performance often degrades. (2) Generation reveals a superior Data Scaling trend and higher Data Utilization. Unified generation-understanding demonstrates a superior scaling trend compared to understanding alone, revealing a more effective way to learn vision-only knowledge directive from vision modality rather than captioning to text. (3) Autoregression on Input Embedding is effective to capture visual details. Compared to the commonly-used vision encoder, make visual autoregression on input embedding shows less cumulative error and is modality independent, which can be extend to all modalities. The learned semantic representations capture visual information such as objects, locations, shapes, and colors; further enable pixel-level image generation.

Zhengkang Guo, Yiyang Li, Lin Qiu et al.

As LLM-based agents increasingly rely on external tools, it is important to evaluate their ability to sustain tool-grounded reasoning beyond familiar workflows and short-range interactions. We introduce AgentEscapeBench, an escape-room-style benchmark that tests whether agents can infer, execute, and revise novel tool-use procedures under explicit long-range dependency constraints. Each task defines a directed acyclic dependency graph over tools and items, requiring agents to invoke real external functions, track hidden state revealed incrementally, propagate intermediate results, and submit a deterministically verifiable final answer. AgentEscapeBench includes 270 instances across five difficulty tiers and supports fully automated evaluation. Experiments with sixteen LLM agents and human participants show that performance drops sharply as dependency depth increases: humans decline from 98.3% success at difficulty-5 to 80.0% at difficulty-25, while the best model drops from 90.0% to 60.0%. Trajectory analysis attributes model failures mainly to breakdowns in long-range state tracking, clue adherence, and intermediate-result propagation. These findings suggest that current agents can often handle local tool use but still struggle with deep contextual dependencies. We hope AgentEscapeBench can serve as a diagnostic testbed for measuring current agent capabilities and informing future training efforts toward more robust general-purpose reasoning, action, and adaptation.

Peng Ding, Jun Kuang, Wen Sun et al.

Large language models (LLMs) remain vulnerable to jailbreaking attacks despite their impressive capabilities. Investigating these weaknesses is crucial for robust safety mechanisms. Existing attacks primarily distract LLMs by introducing additional context or adversarial tokens, leaving the core harmful intent unchanged. In this paper, we introduce ISA (Intent Shift Attack), which obfuscates LLMs about the intent of the attacks. More specifically, we establish a taxonomy of intent transformations and leverage them to generate attacks that may be misperceived by LLMs as benign requests for information. Unlike prior methods relying on complex tokens or lengthy context, our approach only needs minimal edits to the original request, and yields natural, human-readable, and seemingly harmless prompts. Extensive experiments on both open-source and commercial LLMs show that ISA achieves over 70% improvement in attack success rate compared to direct harmful prompts. More critically, fine-tuning models on only benign data reformulated with ISA templates elevates success rates to nearly 100%. For defense, we evaluate existing methods and demonstrate their inadequacy against ISA, while exploring both training-free and training-based mitigation strategies. Our findings reveal fundamental challenges in intent inference for LLMs safety and underscore the need for more effective defenses. Our code and datasets are available at https://github.com/NJUNLP/ISA.

Jiaming Wang, Zhe Tang, Yilin Jin et al.

As large language models (LLMs) are widely deployed as domain-specific agents, many benchmarks have been proposed to evaluate their ability to follow instructions and make decisions in real-world scenarios. However, business scenarios often involve complex standard operating procedures (SOPs), and the evaluation of LLM capabilities in such contexts has not been fully explored. To bridge this gap, we propose SOP-Maze, a benchmark constructed from real-world business data and adapted into a collection of 397 tasks from 23 complex SOP scenarios. We further categorize SOP tasks into two broad classes: Lateral Root System (LRS), representing wide-option tasks that demand precise selection; and Heart Root System (HRS), which emphasizes deep logical reasoning with complex branches. Extensive experiments reveal that nearly all state-of-the-art models struggle with SOP-Maze. We conduct a comprehensive analysis and identify three key error categories: (i) route blindness: difficulty following procedures; (ii) conversational fragility: inability to handle real dialogue nuances; and (iii) calculation errors: mistakes in time or arithmetic reasoning under complex contexts. The systematic study explores LLM performance across SOP tasks that challenge both breadth and depth, offering new insights for improving model capabilities. We have open-sourced our work on https://github.com/ADoublLEN/SOP-Maze.

Zhejian Lai, Xiang Geng, Zhijun Wang et al.

Mathematical reasoning is a primary indicator of large language models (LLMs) intelligence. However, existing LLMs exhibit failures of robustness and generalization. This paper attributes these deficiencies to spurious reasoning, i.e., producing answers from superficial features. To address this challenge, we propose the AdaR framework to enable adaptive reasoning, wherein models rely on problem-solving logic to produce answers. AdaR synthesizes logically equivalent queries by varying variable values, and trains models with RLVR on these data to penalize spurious logic while encouraging adaptive logic. To improve data quality, we extract the problem-solving logic from the original query and generate the corresponding answer by code execution, then apply a sanity check. Experimental results demonstrate that AdaR improves robustness and generalization, achieving substantial improvement in mathematical reasoning while maintaining high data efficiency. Analysis indicates that data synthesis and RLVR function in a coordinated manner to enable adaptive reasoning in LLMs. Subsequent analyses derive key design insights into the effect of critical factors and the applicability to instruct LLMs. Our project is available at https://github.com/NJUNLP/AdaR.

Siyu Yan, Long Zeng, Xuecheng Wu et al.

As large language models~(LLMs) become widely adopted, ensuring their alignment with human values is crucial to prevent jailbreaks where adversaries manipulate models to produce harmful content. While most defenses target single-turn attacks, real-world usage often involves multi-turn dialogues, exposing models to attacks that exploit conversational context to bypass safety measures. We introduce MUSE, a comprehensive framework tackling multi-turn jailbreaks from both attack and defense angles. For attacks, we propose MUSE-A, a method that uses frame semantics and heuristic tree search to explore diverse semantic trajectories. For defense, we present MUSE-D, a fine-grained safety alignment approach that intervenes early in dialogues to reduce vulnerabilities. Extensive experiments on various models show that MUSE effectively identifies and mitigates multi-turn vulnerabilities. Code is available at \href{https://github.com/yansiyu02/MUSE}{https://github.com/yansiyu02/MUSE}.

Ziyan Liu, Junwen Li, Kaiwen Li et al.

Multimodal entity linking plays a crucial role in a wide range of applications. Recent advances in large language model-based methods have become the dominant paradigm for this task, effectively leveraging both textual and visual modalities to enhance performance. Despite their success, these methods still face two challenges, including unnecessary incorporation of image data in certain scenarios and the reliance only on a one-time extraction of visual features, which can undermine their effectiveness and accuracy. To address these challenges, we propose a novel LLM-based framework for the multimodal entity linking task, called Intra- and Inter-modal Collaborative Reflections. This framework prioritizes leveraging text information to address the task. When text alone is insufficient to link the correct entity through intra- and inter-modality evaluations, it employs a multi-round iterative strategy that integrates key visual clues from various aspects of the image to support reasoning and enhance matching accuracy. Extensive experiments on three widely used public datasets demonstrate that our framework consistently outperforms current state-of-the-art methods in the task, achieving improvements of 3.2%, 5.1%, and 1.6%, respectively. Our code is available at https://github.com/ziyan-xiaoyu/I2CR/.

Dan Ma, Jun Xu, Zongyu Wang et al.

Product reviews often contain a large number of implicit aspects and object-attribute co-existence cases. Unfortunately, many existing studies in Aspect-Based Sentiment Analysis (ABSA) have overlooked this issue, which can make it difficult to extract opinions comprehensively and fairly. In this paper, we propose a new task called Entity-Aspect-Opinion-Sentiment Quadruple Extraction (EASQE), which aims to hierarchically decompose aspect terms into entities and aspects to avoid information loss, non-exclusive annotations, and opinion misunderstandings in ABSA tasks. To facilitate research in this new task, we have constructed four datasets (Res14-EASQE, Res15-EASQE, Res16-EASQE, and Lap14-EASQE) based on the SemEval Restaurant and Laptop datasets. We have also proposed a novel two-stage sequence-tagging based Trigger-Opinion framework as the baseline for the EASQE task. Empirical evaluations show that our Trigger-Opinion framework can generate satisfactory EASQE results and can also be applied to other ABSA tasks, significantly outperforming state-of-the-art methods. We have made the four datasets and source code of Trigger-Opinion publicly available to facilitate further research in this area.

Shuhao Han, Haotian Fan, Fangyuan Kong et al.

This paper reports on the NTIRE 2025 challenge on Text to Image (T2I) generation model quality assessment, which will be held in conjunction with the New Trends in Image Restoration and Enhancement Workshop (NTIRE) at CVPR 2025. The aim of this challenge is to address the fine-grained quality assessment of text-to-image generation models. This challenge evaluates text-to-image models from two aspects: image-text alignment and image structural distortion detection, and is divided into the alignment track and the structural track. The alignment track uses the EvalMuse-40K, which contains around 40K AI-Generated Images (AIGIs) generated by 20 popular generative models. The alignment track has a total of 371 registered participants. A total of 1,883 submissions are received in the development phase, and 507 submissions are received in the test phase. Finally, 12 participating teams submitted their models and fact sheets. The structure track uses the EvalMuse-Structure, which contains 10,000 AI-Generated Images (AIGIs) with corresponding structural distortion mask. A total of 211 participants have registered in the structure track. A total of 1155 submissions are received in the development phase, and 487 submissions are received in the test phase. Finally, 8 participating teams submitted their models and fact sheets. Almost all methods have achieved better results than baseline methods, and the winning methods in both tracks have demonstrated superior prediction performance on T2I model quality assessment.

Yingxuan Yang, Bo Huang, Siyuan Qi et al.

Agentic workflows have become the dominant paradigm for building complex AI systems, orchestrating specialized components, such as planning, reasoning, action execution, and reflection, to tackle sophisticated real-world tasks. However, systematically analyzing and optimizing these workflows remains challenging due to intricate component interdependencies and the lack of principled attribution methods. In this work, we introduce ShapleyFlow, the first framework that employs cooperative game theory to analyze and optimize agentic workflows. By applying the Shapley value to evaluate all possible component configurations, ShapleyFlow enables fine-grained attribution of each component's contribution and facilitates the identification of task-specific optimal configurations. Through a constructed dataset evaluated across 7 scenarios, such as navigation, math and OS, we demonstrate 3 key contributions: (1) Theoretical Framework: a principled game-theoretic approach for the attribution of contributions in agentic workflows. (2) Optimal Workflow Discovery: ShapleyFlow identifies task-specific component configurations that consistently outperform workflows relying on a single LLM across all tested tasks. (3) Comprehensive Analysis: we construct and analyze over 1,500 tasks, providing actionable insights and design guidelines for optimizing workflows across multiple domains.

Zijian Zhang, Xuhui Zheng, Xuecheng Wu et al.

While text-to-image (T2I) generation models have achieved remarkable progress in recent years, existing evaluation methodologies for vision-language alignment still struggle with the fine-grained semantic matching. Current approaches based on global similarity metrics often overlook critical token-level correspondences between textual descriptions and visual content. To this end, we present TokenFocus-VQA, a novel evaluation framework that leverages Large Vision-Language Models (LVLMs) through visual question answering (VQA) paradigm with position-specific probability optimization. Our key innovation lies in designing a token-aware loss function that selectively focuses on probability distributions at pre-defined vocabulary positions corresponding to crucial semantic elements, enabling precise measurement of fine-grained semantical alignment. The proposed framework further integrates ensemble learning techniques to aggregate multi-perspective assessments from diverse LVLMs architectures, thereby achieving further performance enhancement. Evaluated on the NTIRE 2025 T2I Quality Assessment Challenge Track 1, our TokenFocus-VQA ranks 2nd place (0.8445, only 0.0001 lower than the 1st method) on public evaluation and 2nd place (0.8426) on the official private test set, demonstrating superiority in capturing nuanced text-image correspondences compared to conventional evaluation methods.

Nan Wang, Yongqi Fan, yansha zhu et al.

Large Language Models (LLMs) face challenges in knowledge-intensive reasoning tasks like classic multi-hop question and answering, which involves reasoning across multiple facts. This difficulty arises because the chain of thoughts (CoTs) generated by LLMs in such tasks often deviate from real or a priori reasoning paths. In contrast, knowledge graphs (KGs) explicitly represent the logical connections between facts through entities and relationships. This reflects a significant gap. Meanwhile, large reasoning models (LRMs), such as o1, have demonstrated that long-step reasoning significantly enhances the performance of LLMs. Building on these insights, we propose KG-o1, a four-stage approach that integrates KGs to enhance the multi-hop reasoning abilities of LLMs. We first filter out initial entities and generate complex subgraphs. Secondly, we construct logical paths for subgraphs and then use knowledge graphs to build a dataset with a complex and extended brainstorming process, which trains LLMs to imitate long-term reasoning. Finally, we employ rejection sampling to generate a self-improving corpus for direct preference optimization (DPO), further refining the LLMs reasoning abilities. We conducted experiments on two simple and two complex datasets. The results show that KG-o1 models exhibit superior performance across all tasks compared to existing LRMs.

Zijian Zhang, Xuecheng Wu, Danlei Huang et al.

Driven by the rapid progress in vision-language models (VLMs), the responsible behavior of large-scale multimodal models has become a prominent research area, particularly focusing on hallucination detection and factuality checking. In this paper, we present the solution for the two tracks of Responsible AI challenge. Inspirations from the general domain demonstrate that a smaller distilled VLM can often outperform a larger VLM that is directly tuned on downstream tasks, while achieving higher efficiency. We thus jointly tackle two tasks from the perspective of knowledge distillation and propose a progressive hybrid knowledge distillation framework termed HKD4VLM. Specifically, the overall framework can be decomposed into Pyramid-like Progressive Online Distillation and Ternary-Coupled Refinement Distillation, hierarchically moving from coarse-grained knowledge alignment to fine-grained refinement. Besides, we further introduce the mapping shift-enhanced inference and diverse augmentation strategies to enhance model performance and robustness. Extensive experimental results demonstrate the effectiveness of our HKD4VLM. Ablation studies provide insights into the critical design choices driving performance gains.

Xuecheng Wu, Jiaxing Liu, Danlei Huang et al.

Visual-Interleaved Chain-of-Thought (VI-CoT) enables MLLMs to continually update their understanding and decisions based on step-wise intermediate visual states (IVS), much like a human would, which demonstrates impressive success in various tasks, thereby leading to emerged advancements in related benchmarks. Despite promising progress, current benchmarks provide models with relatively fixed IVS, rather than free-style IVS, whch might forcibly distort the original thinking trajectories, failing to evaluate their intrinsic reasoning capabilities. More importantly, existing benchmarks neglect to systematically explore the impact factors that IVS would impart to untamed reasoning performance. To tackle above gaps, we introduce a specialized benchmark termed ViC-Bench, consisting of four representive tasks: maze navigation, jigsaw puzzle, embodied long-horizon planning, and complex counting, where each task has dedicated free-style IVS generation pipeline supporting function calls. To systematically examine VI-CoT capability, we propose a thorough evaluation suite incorporating a progressive three-stage strategy with targeted new metrics. Besides, we establish Incremental Prompting Information Injection (IPII) strategy to ablatively explore the prompting factors for VI-CoT. We extensively conduct evaluations for 18 advanced MLLMs, revealing key insights into their VI-CoT capability. Our proposed benchmark is publicly open at Huggingface.

Chen Chen, ZeYang Hu, Fengjiao Chen et al.

Multimodal Large Languages models have been progressing from uni-modal understanding toward unifying visual, audio and language modalities, collectively termed omni models. However, the correlation between uni-modal and omni-modal remains unclear, which requires comprehensive evaluation to drive omni model's intelligence evolution. In this work, we introduce a novel, high-quality, and UNified Omni model benchmark, UNO-Bench. This benchmark is designed to effectively evaluate both UNi-modal and Omni-modal capabilities under a unified ability taxonomy, spanning 44 task types and 5 modality combinations. It includes 1250 human curated samples for omni-modal with 98% cross-modality solvability, and 2480 enhanced uni-modal samples. The human-generated dataset is well-suited to real-world scenarios, particularly within the Chinese context, whereas the automatically compressed dataset offers a 90% increase in speed and maintains 98% consistency across 18 public benchmarks. In addition to traditional multi-choice questions, we propose an innovative multi-step open-ended question format to assess complex reasoning. A general scoring model is incorporated, supporting 6 question types for automated evaluation with 95% accuracy. Experimental result shows the Compositional Law between omni-modal and uni-modal performance and the omni-modal capability manifests as a bottleneck effect on weak models, while exhibiting synergistic promotion on strong models.

Shushi Wang, Zicheng Zhang, Chunyi Li et al.

Quality assessment of AI-generated content is crucial for evaluating model capability and guiding model optimization. However, most existing quality assessment datasets and models provide only a single quality score, which is too coarse to offer targeted guidance for improving generative models. In current applications of AI-generated images, realism and plausibility are two critical dimensions, and with the emergence of unified generation-understanding models, fine-grained evaluation along these dimensions becomes especially effective for improving generative performance. Therefore, we introduce Q-Real, a novel dataset for fine-grained evaluation of realism and plausibility in AI-generated images. Q-Real consists of 3,088 images generated by popular text-to-image models. For each image, we annotate the locations of major entities and provide a set of judgment questions and attribution descriptions for these along the dimensions of realism and plausibility. Considering that recent advances in multi-modal large language models (MLLMs) enable fine-grained evaluation of AI-generated images, we construct Q-Real Bench to evaluate them on two tasks: judgment and grounding with reasoning. Finally, to enhance MLLM capabilities, we design a fine-tuning framework and conduct experiments on multiple MLLMs using our dataset. Experimental results demonstrate the high quality and significance of our dataset and the comprehensiveness of the benchmark. Dataset and code will be released upon publication.

Lingyue Fu, Bolun Zhang, Hao Guan et al.

Recent advances in code agents have enabled automated software development at the project level, supported by large language models (LLMs) and widely adopted tools. However, existing benchmarks for code agent evaluation face two major limitations: high annotation cost and expertise requirements, and rigid evaluation metrics that rely primarily on unit tests. To address these challenges, we propose an agent-driven benchmark construction pipeline that leverages human supervision to efficiently generate diverse and challenging project-level tasks. Based on this approach, we introduce PRDBench, a novel benchmark comprising 50 real-world Python projects across 20 domains, each with structured Product Requirement Document (PRD) requirements, comprehensive evaluation criteria, and reference implementations. PRDBench features rich data sources, high task complexity, and flexible metrics. We further employ an Agent-as-a-Judge paradigm to score agent outputs, enabling the evaluation of various test types beyond unit tests. Extensive experiments on PRDBench demonstrate its effectiveness in assessing the capabilities of both code agents and evaluation agents, providing a scalable and robust framework for annotation and evaluation.

Yi Lu, Jianing Wang, Linsen Guo et al.

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek-R1) have led to remarkable improvements through long Chain-of-Thought (CoT). However, existing benchmarks mainly focus on immediate, single-horizon tasks, failing to adequately evaluate models' ability to understand and respond to complex, long-horizon scenarios. To address this incomplete evaluation of Large Reasoning Models (LRMs), we propose R-HORIZON, a method designed to stimulate long-horizon reasoning behaviors in LRMs through query composition. Based on R-HORIZON, we construct a long-horizon reasoning benchmark, comprising complex multi-step reasoning tasks with interdependent problems that span long reasoning horizons. Through comprehensive evaluation of LRMs using the R-HORIZON benchmark, we find that even the most advanced LRMs suffer significant performance degradation. Our analysis reveals that LRMs exhibit limited effective reasoning length and struggle to allocate thinking budget across multiple problems appropriately. Recognizing these limitations, we use R-HORIZON to construct long-horizon reasoning data for reinforcement learning with verified rewards (RLVR). Compared to training with single-horizon data, RLVR with R-HORIZON not only substantially improves performance on the multi-horizon reasoning tasks, but also promotes accuracy on standard reasoning tasks, with an increase of 7.5 on AIME2024. These results position R-HORIZON as a scalable, controllable, and low-cost paradigm for enhancing and evaluating the long-horizon reasoning capabilities of LRMs.

Yiyang Li, Yonghuang Wu, Ying Luo et al.

Evaluations of large language models (LLMs) suffer from instability, where small changes of random factors such as few-shot examples can lead to drastic fluctuations of scores and even model rankings. Moreover, different LLMs can have different preferences for a certain setting of random factors. As a result, using a fixed setting of random factors, which is often adopted as the paradigm of current evaluations, can lead to potential unfair comparisons between LLMs. To mitigate the volatility of evaluations, we first theoretically analyze the sources of variance induced by changes in random factors. Targeting these specific sources, we then propose the instance-level randomization (ILR) method to reduce variance and enhance fairness in model comparisons. Instead of using a fixed setting across the whole benchmark in a single experiment, we randomize all factors that affect evaluation scores for every single instance, run multiple experiments and report the averaged score. Theoretical analyses and empirical results demonstrate that ILR can reduce the variance and unfair comparisons caused by random factors, as well as achieve similar robustness level with less than half computational cost compared with previous methods.

Zihao Zhao, Jiawei Chen, Sheng Zhou et al.

Recommender system usually suffers from severe popularity bias -- the collected interaction data usually exhibits quite imbalanced or even long-tailed distribution over items. Such skewed distribution may result from the users' conformity to the group, which deviates from reflecting users' true preference. Existing efforts for tackling this issue mainly focus on completely eliminating popularity bias. However, we argue that not all popularity bias is evil. Popularity bias not only results from conformity but also item quality, which is usually ignored by existing methods. Some items exhibit higher popularity as they have intrinsic better property. Blindly removing the popularity bias would lose such important signal, and further deteriorate model performance. To sufficiently exploit such important information for recommendation, it is essential to disentangle the benign popularity bias caused by item quality from the harmful popularity bias caused by conformity. Although important, it is quite challenging as we lack an explicit signal to differentiate the two factors of popularity bias. In this paper, we propose to leverage temporal information as the two factors exhibit quite different patterns along the time: item quality revealing item inherent property is stable and static while conformity that depends on items' recent clicks is highly time-sensitive. Correspondingly, we further propose a novel Time-aware DisEntangled framework (TIDE), where a click is generated from three components namely the static item quality, the dynamic conformity effect, as well as the user-item matching score returned by any recommendation model. Lastly, we conduct interventional inference such that the recommendation can benefit from the benign popularity bias while circumvent the harmful one. Extensive experiments on three real-world datasets demonstrated the effectiveness of TIDE.