Muxi Diao

Runqi Qiao, Qiuna Tan, Guanting Dong et al.

Visual mathematical reasoning, as a fundamental visual reasoning ability, has received widespread attention from the Large Multimodal Models (LMMs) community. Existing benchmarks, such as MathVista and MathVerse, focus more on the result-oriented performance but neglect the underlying principles in knowledge acquisition and generalization. Inspired by human-like mathematical reasoning, we introduce WE-MATH, the first benchmark specifically designed to explore the problem-solving principles beyond end-to-end performance. We meticulously collect and categorize 6.5K visual math problems, spanning 67 hierarchical knowledge concepts and five layers of knowledge granularity. We decompose composite problems into sub-problems according to the required knowledge concepts and introduce a novel four-dimensional metric, namely Insufficient Knowledge (IK), Inadequate Generalization (IG), Complete Mastery (CM), and Rote Memorization (RM), to hierarchically assess inherent issues in LMMs' reasoning process. With WE-MATH, we conduct a thorough evaluation of existing LMMs in visual mathematical reasoning and reveal a negative correlation between solving steps and problem-specific performance. We confirm the IK issue of LMMs can be effectively improved via knowledge augmentation strategies. More notably, the primary challenge of GPT-4o has significantly transitioned from IK to IG, establishing it as the first LMM advancing towards the knowledge generalization stage. In contrast, other LMMs exhibit a marked inclination towards Rote Memorization - they correctly solve composite problems involving multiple knowledge concepts yet fail to answer sub-problems. We anticipate that WE-MATH will open new pathways for advancements in visual mathematical reasoning for LMMs. The WE-MATH data and evaluation code are available at https://github.com/We-Math/We-Math.

Kimi Team, Tongtong Bai, Yifan Bai et al.

We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

Yejie Wang, Keqing He, Dayuan Fu et al.

Recently, there has been a growing interest in studying how to construct better code instruction tuning data. However, we observe Code models trained with these datasets exhibit high performance on HumanEval but perform worse on other benchmarks such as LiveCodeBench. Upon further investigation, we find that many datasets suffer from severe data leakage. After cleaning up most of the leaked data, some well-known high-quality datasets perform poorly. This discovery reveals a new challenge: identifying which dataset genuinely qualify as high-quality code instruction data. To address this, we propose an efficient code data pruning strategy for selecting good samples. Our approach is based on three dimensions: instruction complexity, response quality, and instruction diversity. Based on our selected data, we present XCoder, a family of models finetuned from LLaMA3. Our experiments show XCoder achieves new state-of-the-art performance using fewer training data, which verify the effectiveness of our data strategy. Moreover, we perform a comprehensive analysis on the data composition and find existing code datasets have different characteristics according to their construction methods, which provide new insights for future code LLMs. Our models and dataset are released in https://github.com/banksy23/XCoder

Muxi Diao, Rumei Li, Shiyang Liu et al.

As large language models (LLMs) continue to advance in capability and influence, ensuring their security and preventing harmful outputs has become crucial. A promising approach to address these concerns involves training models to automatically generate adversarial prompts for red teaming. However, the evolving subtlety of vulnerabilities in LLMs challenges the effectiveness of current adversarial methods, which struggle to specifically target and explore the weaknesses of these models. To tackle these challenges, we introduce the $\mathbf{S}\text{elf-}\mathbf{E}\text{volving }\mathbf{A}\text{dversarial }\mathbf{S}\text{afety }\mathbf{(SEAS)}$ optimization framework, which enhances security by leveraging data generated by the model itself. SEAS operates through three iterative stages: Initialization, Attack, and Adversarial Optimization, refining both the Red Team and Target models to improve robustness and safety. This framework reduces reliance on manual testing and significantly enhances the security capabilities of LLMs. Our contributions include a novel adversarial framework, a comprehensive safety dataset, and after three iterations, the Target model achieves a security level comparable to GPT-4, while the Red Team model shows a marked increase in attack success rate (ASR) against advanced models. Our code and datasets are released at https://SEAS-LLM.github.io/.

Muxi Diao, Lele Yang, Wuxuan Gong et al.

Supervised Fine-Tuning (SFT) is the standard paradigm for domain adaptation, yet it frequently incurs the cost of catastrophic forgetting. In sharp contrast, on-policy Reinforcement Learning (RL) effectively preserves general capabilities. We investigate this discrepancy and identify a fundamental distributional gap: while RL aligns with the model's internal belief, SFT forces the model to fit external supervision. This mismatch often manifests as "Confident Conflicts" tokens characterized by low probability but low entropy. In these instances, the model is highly confident in its own prediction but is forced to learn a divergent ground truth, triggering destructive gradient updates. To address this, we propose Entropy-Adaptive Fine-Tuning (EAFT). Unlike methods relying solely on prediction probability, EAFT utilizes token-level entropy as a gating mechanism to distinguish between epistemic uncertainty and knowledge conflict. This allows the model to learn from uncertain samples while suppressing gradients on conflicting data. Extensive experiments on Qwen and GLM series (ranging from 4B to 32B parameters) across mathematical, medical, and agentic domains confirm our hypothesis. EAFT consistently matches the downstream performance of standard SFT while significantly mitigating the degradation of general capabilities.

Xinran Wang, Yuxuan Zhang, Xiao Zhang et al.

Accurately detecting and localizing hallucinations is a critical task for ensuring high reliability of image captions. In the era of Multimodal Large Language Models (MLLMs), captions have evolved from brief sentences into comprehensive narratives, often spanning hundreds of words. This shift exponentially increases the challenge: models must now pinpoint specific erroneous spans or words within extensive contexts, rather than merely flag response-level inconsistencies. However, existing benchmarks lack the fine granularity and domain diversity required to evaluate this capability. To bridge this gap, we introduce DetailVerifyBench, a rigorous benchmark comprising 1,000 high-quality images across five distinct domains. With an average caption length of over 200 words and dense, token-level annotations of multiple hallucination types, it stands as the most challenging benchmark for precise hallucination localization in the field of long image captioning to date. Our benchmark is available at https://zyx-hhnkh.github.io/DetailVerifyBench/.

Zhexu Wang, Yiping Liu, Yejie Wang et al.

Recent advancements in large language models (LLMs) have demonstrated significant progress in math and code reasoning capabilities. However, existing code benchmark are limited in their ability to evaluate the full spectrum of these capabilities, particularly at the competitive level. To bridge this gap, we introduce OJBench, a novel and challenging benchmark designed to assess the competitive-level code reasoning abilities of LLMs. OJBench comprises 232 programming competition problems from NOI and ICPC, providing a more rigorous test of models' reasoning skills. We conducted a comprehensive evaluation using OJBench on 37 models, including both closed-source and open-source models, reasoning-oriented and non-reasoning-oriented models. Our results indicate that even state-of-the-art reasoning-oriented models, such as o4-mini and Gemini-2.5-pro-exp, struggle with highly challenging competition-level problems. This highlights the significant challenges that models face in competitive-level code reasoning.

Xinran Wang, Songyu Xu, Xiangxuan Shan et al.

Cinematography is a cornerstone of film production and appreciation, shaping mood, emotion, and narrative through visual elements such as camera movement, shot composition, and lighting. Despite recent progress in multimodal large language models (MLLMs) and video generation models, the capacity of current models to grasp and reproduce cinematographic techniques remains largely uncharted, hindered by the scarcity of expert-annotated data. To bridge this gap, we present CineTechBench, a pioneering benchmark founded on precise, manual annotation by seasoned cinematography experts across key cinematography dimensions. Our benchmark covers seven essential aspects-shot scale, shot angle, composition, camera movement, lighting, color, and focal length-and includes over 600 annotated movie images and 120 movie clips with clear cinematographic techniques. For the understanding task, we design question answer pairs and annotated descriptions to assess MLLMs' ability to interpret and explain cinematographic techniques. For the generation task, we assess advanced video generation models on their capacity to reconstruct cinema-quality camera movements given conditions such as textual prompts or keyframes. We conduct a large-scale evaluation on 15+ MLLMs and 5+ video generation models. Our results offer insights into the limitations of current models and future directions for cinematography understanding and generation in automatically film production and appreciation. The code and benchmark can be accessed at https://github.com/PRIS-CV/CineTechBench.

Yuxuan Yang, Zhonghao Yan, Yi Zhang et al.

Hepatocellular Carcinoma diagnosis relies heavily on the interpretation of gigapixel Whole Slide Images. However, current computational approaches are constrained by fixed-resolution processing mechanisms and inefficient feature aggregation, which inevitably lead to either severe information loss or high feature redundancy. To address these challenges, we propose Hepato-LLaVA, a specialized Multi-modal Large Language Model designed for fine-grained hepatocellular pathology analysis. We introduce a novel Sparse Topo-Pack Attention mechanism that explicitly models 2D tissue topology. This mechanism effectively aggregates local diagnostic evidence into semantic summary tokens while preserving global context. Furthermore, to overcome the lack of multi-scale data, we present HepatoPathoVQA, a clinically grounded dataset comprising 33K hierarchically structured question-answer pairs validated by expert pathologists. Our experiments demonstrate that Hepato-LLaVA achieves state-of-the-art performance on HCC diagnosis and captioning tasks, significantly outperforming existing methods. Our code and implementation details are available at https://pris-cv.github.io/Hepto-LLaVA/.

Muxi Diao, Yutao Mou, Keqing He et al.

The safety of Large Language Models (LLMs) is crucial for the development of trustworthy AI applications. Existing red teaming methods often rely on seed instructions, which limits the semantic diversity of the synthesized adversarial prompts. We propose AutoRed, a free-form adversarial prompt generation framework that removes the need for seed instructions. AutoRed operates in two stages: (1) persona-guided adversarial instruction generation, and (2) a reflection loop to iteratively refine low-quality prompts. To improve efficiency, we introduce a verifier to assess prompt harmfulness without querying the target models. Using AutoRed, we build two red teaming datasets -- AutoRed-Medium and AutoRed-Hard -- and evaluate eight state-of-the-art LLMs. AutoRed achieves higher attack success rates and better generalization than existing baselines. Our results highlight the limitations of seed-based approaches and demonstrate the potential of free-form red teaming for LLM safety evaluation. We will open source our datasets in the near future.

Xiaoshuai Song, Muxi Diao, Guanting Dong et al.

Large language models (LLMs) have demonstrated significant potential in advancing various fields of research and society. However, the current community of LLMs overly focuses on benchmarks for analyzing specific foundational skills (e.g. mathematics and code generation), neglecting an all-round evaluation of the computer science field. To bridge this gap, we introduce CS-Bench, the first multilingual (English, Chinese, French, German) benchmark dedicated to evaluating the performance of LLMs in computer science. CS-Bench comprises approximately 10K meticulously curated test samples, covering 26 subfields across 4 key areas of computer science, encompassing various task forms and divisions of knowledge and reasoning. Utilizing CS-Bench, we conduct a comprehensive evaluation of over 30 mainstream LLMs, revealing the relationship between CS performance and model scales. We also quantitatively analyze the reasons for failures in existing LLMs and highlight directions for improvements, including knowledge supplementation and CS-specific reasoning. Further cross-capability experiments show a high correlation between LLMs' capabilities in computer science and their abilities in mathematics and coding. Moreover, expert LLMs specialized in mathematics and coding also demonstrate strong performances in several CS subfields. Looking ahead, we envision CS-Bench serving as a cornerstone for LLM applications in the CS field and paving new avenues in assessing LLMs' diverse reasoning capabilities. The CS-Bench data and evaluation code are available at https://github.com/csbench/csbench.

Xinran Wang, Muxi Diao, Baoteng Li et al.

The Controllable Image Captioning Agent (CapAgent) is an innovative system designed to bridge the gap between user simplicity and professional-level outputs in image captioning tasks. CapAgent automatically transforms user-provided simple instructions into detailed, professional instructions, enabling precise and context-aware caption generation. By leveraging multimodal large language models (MLLMs) and external tools such as object detection tool and search engines, the system ensures that captions adhere to specified guidelines, including sentiment, keywords, focus, and formatting. CapAgent transparently controls each step of the captioning process, and showcases its reasoning and tool usage at every step, fostering user trust and engagement. The project code is available at https://github.com/xin-ran-w/CapAgent.

Yejie Wang, Keqing He, Guanting Dong et al.

Code Large Language Models (Code LLMs) have demonstrated outstanding performance in code-related tasks. Several instruction tuning approaches have been proposed to boost the code generation performance of pre-trained Code LLMs. In this paper, we introduce a diverse instruction model (DolphCoder) with self-evaluating for code generation. It learns diverse instruction targets and combines a code evaluation objective to enhance its code generation ability. Our model achieves superior performance on the HumanEval and MBPP benchmarks, demonstrating new insights for future code instruction tuning work. Our key findings are: (1) Augmenting more diverse responses with distinct reasoning paths increases the code capability of LLMs. (2) Improving one's ability to evaluate the correctness of code solutions also enhances their ability to create it.

Yueying Gao, Dongliang Chang, Bingyao Yu et al.

Accurate and interpretable detection of AI-generated images is essential for mitigating risks associated with AI misuse. However, the substantial domain gap among generative models makes it challenging to develop a generalizable forgery detection model. Moreover, since every pixel in an AI-generated image is synthesized, traditional saliency-based forgery explanation methods are not well suited for this task. To address these challenges, we formulate detection and explanation as a unified Forgery Detection and Reasoning task (FDR-Task), leveraging Multi-Modal Large Language Models (MLLMs) to provide accurate detection through reliable reasoning over forgery attributes. To facilitate this task, we introduce the Multi-Modal Forgery Reasoning dataset (MMFR-Dataset), a large-scale dataset containing 120K images across 10 generative models, with 378K reasoning annotations on forgery attributes, enabling comprehensive evaluation of the FDR-Task. Furthermore, we propose FakeReasoning, a forgery detection and reasoning framework with three key components: 1) a dual-branch visual encoder that integrates CLIP and DINO to capture both high-level semantics and low-level artifacts; 2) a Forgery-Aware Feature Fusion Module that leverages DINO's attention maps and cross-attention mechanisms to guide MLLMs toward forgery-related clues; 3) a Classification Probability Mapper that couples language modeling and forgery detection, enhancing overall performance. Experiments across multiple generative models demonstrate that FakeReasoning not only achieves robust generalization but also outperforms state-of-the-art methods on both detection and reasoning tasks.

Muxi Diao, Lele Yang, Hongbo Yin et al.

Autonomous driving requires real-time, robust reasoning across perception, prediction, planning, and behavior. However, conventional end-to-end models fail to generalize in complex scenarios due to the lack of structured reasoning. Recent vision-language models (VLMs) have been applied to driving tasks, but they typically rely on isolated modules and static supervision, limiting their ability to support multi-stage decision-making. We present AutoDriveRL, a unified training framework that formulates autonomous driving as a structured reasoning process over four core tasks. Each task is independently modeled as a vision-language question-answering problem and optimized using task-specific reward models, enabling fine-grained reinforcement signals at different reasoning stages. Within this framework, we train DriveRX, a cross-task reasoning VLM designed for real-time decision-making. DriveRX achieves strong performance on a public benchmark, outperforming GPT-4o in behavior reasoning and demonstrating robustness under complex or corrupted driving conditions. Our analysis further highlights the impact of vision encoder design and reward-guided reasoning compression. We will release the AutoDriveRL framework and the DriveRX model to support future research.

Zhonghao Yan, Muxi Diao, Yuxuan Yang et al.

Accurately grounding regions of interest (ROIs) is critical for diagnosis and treatment planning in medical imaging. While multimodal large language models (MLLMs) combine visual perception with natural language, current medical-grounding pipelines still rely on supervised fine-tuning with explicit spatial hints, making them ill-equipped to handle the implicit queries common in clinical practice. This work makes three core contributions. We first define Unified Medical Reasoning Grounding (UMRG), a novel vision-language task that demands clinical reasoning and pixel-level grounding. Second, we release U-MRG-14K, a dataset of 14K samples featuring pixel-level masks alongside implicit clinical queries and reasoning traces, spanning 10 modalities, 15 super-categories, and 108 specific categories. Finally, we introduce MedReasoner, a modular framework that distinctly separates reasoning from segmentation: an MLLM reasoner is optimized with reinforcement learning, while a frozen segmentation expert converts spatial prompts into masks, with alignment achieved through format and accuracy rewards. MedReasoner achieves state-of-the-art performance on U-MRG-14K and demonstrates strong generalization to unseen clinical queries, underscoring the significant promise of reinforcement learning for interpretable medical grounding.

Zixi Yang, Jiapeng Li, Muxi Diao et al.

Recently, Multi-modal Large Language Models (MLLMs) have demonstrated significant performance across various video understanding tasks. However, their robustness, particularly when faced with manipulated video content, remains largely unexplored. In this paper, we introduce Ro-Bench, the first benchmark for evaluating MLLMs on dynamic out-of-distribution (OOD) counterfactual video test sets. Ro-Bench incorporates high-quality, diverse and temporally relevant video data, by editing Style, Object, Background and their compositions. We evaluated eight recent video MLLMs and found that current models exhibit substantial performance degradation on Ro-Bench when exposed to counterfactual video content. Furthermore, we demonstrate that fine-tuning MLLMs with counterfactual data enhances robustness, achieving a 21.73% performance increase on Ro-Bench and a 12.78% improvement across 20 tasks in the MVBench dataset. These findings underscore the effectiveness of counterfactual data in enhancing the video understanding ability of MLLMs. The code and data will be released shortly.

Xinran Wang, Muxi Diao, Yuanzhi Liu et al.

Training text-to-image (T2I) models with detailed captions can significantly improve their generation quality. Existing methods often rely on simplistic metrics like caption length to represent the detailness of the caption in the T2I training set. In this paper, we propose a new metric to estimate caption detailness based on two aspects: image coverage rate (ICR), which evaluates whether the caption covers all regions/objects in the image, and average object detailness (AOD), which quantifies the detailness of each object's description. Through experiments on the COCO dataset using ShareGPT4V captions, we demonstrate that T2I models trained on high-ICR and -AOD captions achieve superior performance on DPG and other benchmarks. Notably, our metric enables more effective data selection-training on only 20% of full data surpasses both full-dataset training and length-based selection method, improving alignment and reconstruction ability. These findings highlight the critical role of detail-aware metrics over length-based heuristics in caption selection for T2I tasks.

Pei Wang, Yejie Wang, Muxi Diao et al.

In the deployment of large language models (LLMs), accurate confidence estimation is critical for assessing the credibility of model predictions. However, existing methods often fail to overcome the issue of overconfidence on incorrect answers. In this work, we focus on improving the confidence estimation of large language models. Considering the fragility of self-awareness in language models, we introduce a Multi-Perspective Consistency (MPC) method. We leverage complementary insights from different perspectives within models (MPC-Internal) and across different models (MPC-Across) to mitigate the issue of overconfidence arising from a singular viewpoint. The experimental results on eight publicly available datasets show that our MPC achieves state-of-the-art performance. Further analyses indicate that MPC can mitigate the problem of overconfidence and is effectively scalable to other models.