Zhitao He

Feng Ju, Zeyu Qin, Rui Min et al.

While Test-Time Scaling (TTS) has proven effective in improving the reasoning ability of large language models (LLMs), low diversity in model outputs often becomes a bottleneck; this is partly caused by the common "one problem, one solution" (1P1S) training practice, which provides a single canonical answer and can push models toward a narrow set of reasoning paths. To address this, we propose a "one problem, multiple solutions" (1PNS) training paradigm that exposes the model to a variety of valid reasoning trajectories and thus increases inference diversity. A core challenge for 1PNS is reliably measuring semantic differences between multi-step chains of thought, so we introduce Reasoning Path Divergence (RPD), a step-level metric that aligns and scores Long Chain-of-Thought solutions to capture differences in intermediate reasoning. Using RPD, we curate maximally diverse solution sets per problem and fine-tune Qwen3-4B-Base. Experiments show that RPD-selected training yields more varied outputs and higher pass@k, with an average +2.80% gain in pass@16 over a strong 1P1S baseline and a +4.99% gain on AIME24, demonstrating that 1PNS further amplifies the effectiveness of TTS. Our code is available at https://github.com/fengjujf/Reasoning-Path-Divergence .

MiniMax, Aili Chen, Aonian Li et al.

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

Haolin Yang, Jipeng Zhang, Zhitao He et al.

Translating natural language to SQL remains difficult for complex queries. Such queries often need environmental interaction and self-correction. To address this, we introduce MARS-SQL, a novel multi-agent framework that combines principled task decomposition and interactive reinforcement learning (RL). Our system comprises three specialized agents: a Grounding Agent for schema linking, a Generation Agent for query generation, and a Validation Agent for final selection. The core of our framework is the Generation agent, which is trained via a multi-turn RL policy. Adopting a ReAct-style Think-Act-Observe loop, the agent iteratively generates thoughts, executes SQL actions against a live database, and revises its strategy based on execution feedback, enabling dynamic, stateful reasoning and self-correction. At inference time, we generate multiple interaction trajectories to explore diverse reasoning paths. The Validation agent, then selects the optimal trajectory by modeling verification as a next-token prediction task and choosing the solution with the highest generation probability. This structured workflow pipelines specialized agents. It combines interactive RL for generation with generative modeling for verification. The approach proves highly effective for robust and accurate SQL generation. Experiments show that MARS-SQL achieves state-of-the-art Execution Accuracy of 77.84% on the BIRD dev set and 89.75% on the Spider test set. Our code is available at https://github.com/YangHaolin0526/MARS-SQL.

Zhitao He, Haolin Yang, Rui Min et al.

Large Language Models (LLMs) excel at long-context understanding but exhibit significant limitations in long-form generation. Existing studies primarily focus on single-generation quality, generally overlooking the volatility of the output. This volatility not only leads to significant computational costs but also severely impacts the models' reliable application. To address this gap, our work unfolds in three stages: benchmarking, probing, and mitigation. We first propose the VOlatility in Long-form Text Benchmark (VOLTBench), a novel heterogeneous-task benchmark designed to systematically quantify the length volatility of long-form generation. Subsequently, by analyzing attention traces, we conduct an in-depth probe to identify several common internal patterns that cause this volatility. Finally, to mitigate long-form output volatility, we propose Stable Generation via Logits Boosting (GLoBo), a lightweight decoding-stage optimization strategy, designed to significantly enhance both the length accuracy and stability of long-form generation without additional training. Extensive experiments on VOLTBench provide the first systematic confirmation of severe long-form output instability in mainstream models and validate that our proposed method successfully improves the mean output length of the base model by 148% and reduces the length volatility by 69%, while maintaining high generation quality.

Yuxin Li, Minghao Liu, Ruida Wang et al.

We present **Lean4PHYS**, a comprehensive reasoning framework for college-level physics problems in Lean4. **Lean4PHYS** includes *LeanPhysBench*, a college-level benchmark for formal physics reasoning in Lean4, which contains 200 hand-crafted and peer-reviewed statements derived from university textbooks and physics competition problems. To establish a solid foundation for formal reasoning in physics, we also introduce *PhysLib*, a community-driven repository containing fundamental unit systems and theorems essential for formal physics reasoning. Based on the benchmark and Lean4 repository we composed in **Lean4PHYS**, we report baseline results using major expert Math Lean4 provers and state-of-the-art closed-source models, with the best performance of DeepSeek-Prover-V2-7B achieving only 16% and Claude-Sonnet-4 achieving 35%. We also conduct a detailed analysis showing that our *PhysLib* can achieve an average improvement of 11.75% in model performance. This demonstrates the challenging nature of our *LeanPhysBench* and the effectiveness of *PhysLib*. To the best of our knowledge, this is the first study to provide a physics benchmark in Lean4.

Yuchen Huang, Zhiyuan Fan, Zhitao He et al.

Pretrained vision-language models (VLMs) such as CLIP excel in general multimodal comprehension but often struggle to capture nuanced, context-dependent visual cues. This makes it difficult to distinguish between similar-looking concepts with potentially different cultural meanings. Such deficiencies are mainly due to a limited amount of high-quality cultural data, contextual information, and the lack of negative examples that highlight subtle differences. To mitigate this, we design a data curation pipeline leveraging open-sourced VLMs and text-to-image models to construct CulTwin, a synthetic cultural dataset. This dataset consists of paired concept-caption-image triplets, where concepts visually resemble each other but are culturally different. Then, we fine-tune CLIP on CulTwin to develop CultureCLIP, which aligns cultural concepts with contextually enhanced captions and synthetic images through tailored contrastive learning. Experiments on culture-specific benchmarks show that CultureCLIP outperforms the base CLIP, achieving up to a notable 5.49% improvement in fine-grained concept recognition on certain tasks while preserving CLIP's original generalization ability, validating the effectiveness of our data synthesis and VLM backbone training paradigm in capturing subtle cultural distinctions.

Zhitao He, Zongwei Lyu, Yi R Fung

Although artificial intelligence (AI) has become deeply integrated into various stages of the research workflow and achieved remarkable advancements, academic rebuttal remains a significant and underexplored challenge. This is because rebuttal is a complex process of strategic communication under severe information asymmetry rather than a simple technical debate. Consequently, current approaches struggle as they largely imitate surface-level linguistics, missing the essential element of perspective-taking required for effective persuasion. In this paper, we introduce RebuttalAgent, the first framework to ground academic rebuttal in Theory of Mind (ToM), operationalized through a ToM-Strategy-Response (TSR) framework that models reviewer mental state, formulates persuasion strategy, and generates evidence-based response. To train our agent, we construct RebuttalBench, a large-scale dataset synthesized via a novel critique-and-refine approach. Our training process consists of two stages, beginning with a supervised fine-tuning phase to equip the agent with ToM-based analysis and strategic planning capabilities, followed by a reinforcement learning phase leveraging the self-reward mechanism for scalable self-improvement. For reliable and efficient automated evaluation, we further develop Rebuttal-RM, a specialized evaluator trained on over 100K samples of multi-source rebuttal data, which achieves scoring consistency with human preferences surpassing powerful judge GPT-4.1. Extensive experiments show RebuttalAgent significantly outperforms the base model by an average of 18.3% on automated metrics, while also outperforming advanced proprietary models across both automated and human evaluations.

Yixiang Chen, Tianshi Zheng, Shijue Huang et al.

Test-time scaling without interpreter feedback is essential for real-world code generation scenarios where test cases are not readily available. While existing paradigms often rely on either greedy exploitation (i.e., iterative refinement) or stochastic exploration (i.e., relying on sample-based voting or reranking mechanisms), the balance between these two dimensions remains underexplored. To investigate the LLM's intrinsic ability to balance exploitation and exploration, we introduce SELF-REDRAFT, a framework built upon Self-Refine that encourages the model to propose new drafts for solutions that are fundamentally flawed. Our results show that SELF-REDRAFT consistently achieves better performance than Self-Refine when converged under the same maximum number of iterations. Still, we observe that significant room for improvement remains, largely due to two core aspects of current self-redraft capabilities: constrained capacity for generating instructive feedback and fragile discriminative judgment. We also find that balancing strategies vary notably across different LLMs, reflecting distinct, model-specific behaviors. Overall, our study establishes a baseline for intrinsic exploration-exploitation balancing in test-time scaling and identifies feedback and discrimination as key areas with potential for future advances.

Zhitao He, Pengfei Cao, Chenhao Wang et al.

With the development of deep learning, natural language processing technology has effectively improved the efficiency of various aspects of the traditional judicial industry. However, most current efforts focus on tasks within individual judicial stages, making it difficult to handle complex tasks that span multiple stages. As the autonomous agents powered by large language models are becoming increasingly smart and able to make complex decisions in real-world settings, offering new insights for judicial intelligence. In this paper, (1) we propose a novel multi-agent framework, AgentsCourt, for judicial decision-making. Our framework follows the classic court trial process, consisting of court debate simulation, legal resources retrieval and decision-making refinement to simulate the decision-making of judge. (2) we introduce SimuCourt, a judicial benchmark that encompasses 420 Chinese judgment documents, spanning the three most common types of judicial cases. Furthermore, to support this task, we construct a large-scale legal knowledge base, Legal-KB, with multi-resource legal knowledge. (3) Extensive experiments show that our framework outperforms the existing advanced methods in various aspects, especially in generating legal articles, where our model achieves significant improvements of 8.6% and 9.1% F1 score in the first and second instance settings, respectively.

Dadi Guo, Jiayu Liu, Zhiyuan Fan et al.

Large reasoning models (e.g., R1, o3) have demonstrated remarkable mathematical problem-solving abilities. However, the high reported accuracy of these advanced models on popular datasets, reliance on purely numerical evaluation and potential benchmark leakage, often masks their true reasoning shortcomings. To address this, we propose leveraging the inherent rigor and methodological complexity of mathematical proofs as a diagnostic tool to expose these hidden failures. Specifically, we introduce the RFMDataset (Reveal Failure Modes), a collection of 200 diverse mathematical proof problems, and thoroughly evaluate advanced models' performance on it. Our in-depth analysis of their failures uncovers 10 fine-grained error types, which shows fundamental limitations in current large reasoning models: 1) large reasoning models grapple profoundly with mathematical proofs, with some generating entirely correct proofs for less than 20% of problems and failing even on basic ones; 2) models exhibit a diverse spectrum of reasoning failures, prominently demonstrating the lack of guarantees for the correctness and rigor of single-step reasoning; and 3) models show hallucination and incompleteness during the reasoning process. Our findings reveal that models' self-reflection is insufficient to resolve the current logical dilemmas, necessitating formalized and fine-grained logical training.

Zhitao He, Sandeep Polisetty, Zhiyuan Fan et al.

In recent years, multimodal large language models (MLLMs) have made significant progress but continue to face inherent challenges in multimodal reasoning, which requires multi-level (e.g., perception, reasoning) and multi-granular (e.g., multi-step reasoning chain) advanced inferencing. Prior work on estimating model confidence tends to focus on the overall response for training and calibration, but fails to assess confidence in each reasoning step, leading to undesirable hallucination snowballing. In this work, we present MMBoundary, a novel framework that advances the knowledge boundary awareness of MLLMs through reasoning step confidence calibration. To achieve this, we propose to incorporate complementary textual and cross-modal self-rewarding signals to estimate confidence at each step of the MLLM reasoning process. In addition to supervised fine-tuning MLLM on this set of self-rewarded confidence estimation signal for initial confidence expression warm-up, we introduce a reinforcement learning stage with multiple reward functions for further aligning model knowledge and calibrating confidence at each reasoning step, enhancing reasoning chain self-correction. Empirical results show that MMBoundary significantly outperforms existing methods across diverse domain datasets and metrics, achieving an average of 7.5% reduction in multimodal confidence calibration errors and up to 8.3% improvement in task performance.

Zhitao He, Zijun Liu, Peng Li et al. · tsinghua

LLM-based agents have made significant advancements in interactive environments, such as mobile operations and web browsing, and other domains beyond computer using. Current multi-agent systems universally excel in performance, compared to single agents, but struggle with generalization across environments due to predefined roles and inadequate strategies for generalizing language agents. The challenge of achieving both strong performance and good generalization has hindered the progress of multi-agent systems for interactive environments. To address these issues, we propose CollabUIAgents, a multi-agent reinforcement learning framework with a novel multi-agent credit re-assignment (CR) strategy, assigning process rewards with LLMs rather than environment-specific rewards and learning with synthesized preference data, in order to foster generalizable, collaborative behaviors among the role-free agents' policies. Empirical results show that our framework improves both performance and cross-environment generalizability of multi-agent systems. Moreover, our 7B-parameter system achieves results on par with or exceed strong closed-source models, and the LLM that guides the CR. We also provide insights in using granular CR rewards effectively for environment generalization, and accommodating trained LLMs in multi-agent systems.

Zhitao He, Pengfei Cao, Zhuoran Jin et al.

Event Causality Identification (ECI) refers to the detection of causal relations between events in texts. However, most existing studies focus on sentence-level ECI with high-resource languages, leaving more challenging document-level ECI (DECI) with low-resource languages under-explored. In this paper, we propose a Heterogeneous Graph Interaction Model with Multi-granularity Contrastive Transfer Learning (GIMC) for zero-shot cross-lingual document-level ECI. Specifically, we introduce a heterogeneous graph interaction network to model the long-distance dependencies between events that are scattered over a document. Then, to improve cross-lingual transferability of causal knowledge learned from the source language, we propose a multi-granularity contrastive transfer learning module to align the causal representations across languages. Extensive experiments show our framework outperforms the previous state-of-the-art model by 9.4% and 8.2% of average F1 score on monolingual and multilingual scenarios respectively. Notably, in the multilingual scenario, our zero-shot framework even exceeds GPT-3.5 with few-shot learning by 24.3% in overall performance.

Junsheng Huang, Zhitao He, Yucheng Huang et al.

The hallucination of non-existent facts by LLMs is an important problem given its widespread adoption across various applications. Previous research addresses this problem by analyzing the internal parameterized knowledge boundaries to estimate confidence. However, these studies focus on the single-problem setting and have not explored the more challenging multi-problem setting, which requires accurately answering multiple questions simultaneously. We introduce a novel method for the multi-problem setting, Multiple Answers and Confidence Stepwise Tuning (MAC-Tuning), that separates the learning of answer prediction and confidence estimation during fine-tuning on instruction data. Extensive experiments demonstrate that our method outperforms baselines by up to 25\% in average precision.

Minghao Liu, Zhitao He, Zhiyuan Fan et al.

Text-guided image editing has seen significant progress in natural image domains, but its application in medical imaging remains limited and lacks standardized evaluation frameworks. Such editing could revolutionize clinical practices by enabling personalized surgical planning, enhancing medical education, and improving patient communication. To bridge this gap, we introduce MedEBench1, a robust benchmark designed to diagnose reliability in text-guided medical image editing. MedEBench consists of 1,182 clinically curated image-prompt pairs covering 70 distinct editing tasks and 13 anatomical regions. It contributes in three key areas: (1) a clinically grounded evaluation framework that measures Editing Accuracy, Context Preservation, and Visual Quality, complemented by detailed descriptions of intended edits and corresponding Region-of-Interest (ROI) masks; (2) a comprehensive comparison of seven state-of-theart models, revealing consistent patterns of failure; and (3) a diagnostic error analysis technique that leverages attention alignment, using Intersection-over-Union (IoU) between model attention maps and ROI masks to identify mislocalization issues, where models erroneously focus on incorrect anatomical regions. MedEBench sets the stage for developing more reliable and clinically effective text-guided medical image editing tools.

Zhitao He, Haolin Yang, Zeyu Qin et al.

The significant gap between rising demands for clinical training and the scarcity of expert instruction poses a major challenge to medical education. With powerful capabilities in personalized guidance, Large Language Models (LLMs) offer a promising solution to bridge this gap. However, current research focuses mainly on one-on-one knowledge instruction, overlooking collaborative reasoning, a key skill for students developed in teamwork like ward rounds. To this end, we develop ClinEdu, a multi-agent pedagogical simulator with personality-driven patients and diverse student cohorts, enabling controlled testing of complex pedagogical processes and scalable generation of teaching data. Based on ClinEdu, we construct ClinTeach, a large Socratic teaching dialogue dataset that captures the complexities of group instruction. We then train MedTutor-R1, the first multimodal Socratic tutor designed for one-to-many instruction in clinical medical education. MedTutor-R1 is first instruction-tuned on our ClinTeach dataset and then optimized with reinforcement learning, using rewards derived from a three-axis rubric, covering structural fidelity, analytical quality, and clinical safety, to refine its adaptive Socratic strategies. For authentic in-situ assessment, we use simulation-based interactive evaluation that redeploys the tutor back into ClinEdu. Experimental results demonstrate that our MedTutor-R1 outperforms the base model by over 20% in average pedagogical score and is comparable to o3, while also exhibiting high adaptability in handling a varying number of students. This promising performance underscores the effectiveness of our pedagogical simulator, ClinEdu.

Zhuoran Jin, Pengfei Cao, Chenhao Wang et al.

Large language models (LLMs) inevitably memorize sensitive, copyrighted, and harmful knowledge from the training corpus; therefore, it is crucial to erase this knowledge from the models. Machine unlearning is a promising solution for efficiently removing specific knowledge by post hoc modifying models. In this paper, we propose a Real-World Knowledge Unlearning benchmark (RWKU) for LLM unlearning. RWKU is designed based on the following three key factors: (1) For the task setting, we consider a more practical and challenging unlearning setting, where neither the forget corpus nor the retain corpus is accessible. (2) For the knowledge source, we choose 200 real-world famous people as the unlearning targets and show that such popular knowledge is widely present in various LLMs. (3) For the evaluation framework, we design the forget set and the retain set to evaluate the model's capabilities across various real-world applications. Regarding the forget set, we provide four four membership inference attack (MIA) methods and nine kinds of adversarial attack probes to rigorously test unlearning efficacy. Regarding the retain set, we assess locality and utility in terms of neighbor perturbation, general ability, reasoning ability, truthfulness, factuality, and fluency. We conduct extensive experiments across two unlearning scenarios, two models and six baseline methods and obtain some meaningful findings. We release our benchmark and code publicly at http://rwku-bench.github.io for future work.