# Chaofan Tao (陶超凡): full publication list > AI researcher at LegoX and Huawei Foundation Model Lab working on LLM coding agents, agentic RL, agentic data and environments, and efficient LLM training and inference. PhD, The University of Hong Kong. ACL 2022 Outstanding Paper Award. Homepage: https://chaofantao.top/ · Summary: https://chaofantao.top/llms.txt ## Agentic Coding, Reasoning and Evaluation ### REAgent: Requirement-Driven LLM Agents for Software Issue Resolution - Venue: ICSE 2027 - Authors: Shiqi Kuang, Zhao Tian, Kaiwei Lin, Chaofan Tao, Shaowei Wang, Haoli Bai, Lifeng Shang, Junjie Chen - Paper: https://arxiv.org/abs/2604.06861 REAgent turns issue descriptions into structured, information-rich requirements, detects and iteratively refines low-quality requirements, and uses the refined specifications to guide more accurate software patches. 中文摘要：REAgent将软件问题描述转化为结构化、信息丰富的需求规范，识别并迭代完善低质量需求，以引导大语言模型生成更准确的软件补丁。 ### SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving - Venue: arXiv 2026 - Authors: Chaofan Tao, Jierun Chen, Yuxin Jiang, Kaiqi Kou, Shaowei Wang, Ruoyu Wang, Xiaohui Li, Sidi Yang, Yiming Du, Jianbo Dai, Zhiming Mao, Xinyu Wang, Lifeng Shang, Haoli Bai - Paper: https://arxiv.org/abs/2601.01426 - Code: https://github.com/LegoX/SWE-Lego SWE-Lego presents a strong SFT-only recipe (new dataset + curriculum/error masking) and studies verifier-based test-time scaling, pushing open 8B/32B models to SOTA on SWE-bench Verified. 中文摘要：SWE-Lego提出仅用监督微调的软工代码智能体配方（真实+合成数据集+课程学习/错误掩码），并研究了软工场景的测试时扩展，将开源8B/32B模型在SWE-bench Verified上推至SOTA。 ### What Makes Interaction Trajectories Effective for Training Terminal Agents? - Venue: arXiv 2026 - Authors: Sidi Yang, Chaofan Tao, Jierun Chen, Tiezheng Yu, Ruoyu Wang, Yuxin Jiang, Yiming Du, Wendong Xu, Jing Xiong, Taiqiang Wu, Lifeng Shang, Xiaohui Li, Ngai Wong, Haoli Bai - Paper: https://arxiv.org/abs/2606.03461 - Code: https://github.com/SWE-Lego/terminal-lego - Data: https://huggingface.co/datasets/SWE-Lego/Terminal-Lego-15k Using Terminal-Lego, a pipeline that turns real-world issues into environment-verified terminal tasks, the paper finds a pedagogical paradox: stronger agents are not better teachers. Environment-grounded inspect-act-verify trajectories transfer best, letting Qwen3-32B reach 24.3% on Terminal-Bench 2.0 with only 15.3k trajectories, about 30× less data than prior results. 中文摘要：基于将真实问题转化为环境可验证终端任务的Terminal-Lego流水线，论文发现“教学悖论”：更强的智能体不一定是更好的教师。显式呈现“观察-行动-验证”的环境锚定轨迹迁移效果最好，仅用1.53万条轨迹即让Qwen3-32B在Terminal-Bench 2.0上达到24.3%，所需数据约为此前结果的1/30。 ### SwingArena: Competitive Programming Arena for Long-context GitHub Issue Solving - Venue: ICLR · Oral 2026 - Authors: Wendong Xu, Jing Xiong, Chenyang Zhao, Qiujiang Chen, Haoran Wang, Hui Shen, Zhongwei Wan, Jianbo Dai, Taiqiang Wu, He Xiao, Chaofan Tao, Z Morley Mao, Ying Sheng, Zhijiang Guo, Hongxia Yang, Bei Yu, Lingpeng Kong, Quanquan Gu, Ngai Wong - Paper: https://arxiv.org/abs/2505.23932 - Code: https://github.com/menik1126/Swing-Bench SwingArena is a CI-driven competitive benchmark where LLMs take turns as submitters and reviewers, paired with retrieval-augmented code generation to handle long-context, multi-language GitHub issue solving. 中文摘要：SwingArena是一个CI驱动的测评框架，让大语言模型扮演代码的提交者和审阅者角色，配合检索增强的代码生成来处理长上下文、多编程语言的GitHub问题解决。 ### MMSearch-Plus: Benchmarking Provenance-Aware Search for Multimodal Browsing Agents - Venue: ICLR 2026 - Authors: Xijia Tao, Yihua Teng, Xinxing Su, Xinyu Fu, Jihao Wu, Chaofan Tao, Ziru Liu, Haoli Bai, Rui Liu, Lingpeng Kong - Paper: https://arxiv.org/abs/2508.21475 - Repository: https://github.com/mmsearch-plus/MMSearch-Plus MMSearch-Plus is a 311-task benchmark and agent framework with set-of-mark tools that require provenance-aware, vision-in-the-loop search and cross-validation under retrieval noise. 中文摘要：MMSearch-Plus是一个包含311个任务的benchmark和多模态智能体搜索问答框架，配备set-of-mark工具，要求在检索噪声下进行可追溯、引入视觉搜索和交叉验证。 ### OVD: On-policy Verbal Distillation - Venue: arXiv 2026 - Authors: Jing Xiong, Hui Shen, Shansan Gong, Yuxin Cheng, Jianghan Shen, Chaofan Tao, Haochen Tan, Haoli Bai, Lifeng Shang, Ngai Wong - Paper: https://arxiv.org/abs/2601.21968 - Code: https://github.com/menik1126/OVD-code On-policy Verbal Distillation replaces token-level matching with trajectory-level distillation using discrete teacher scores, reducing memory and improving student performance on web QA and math reasoning. 中文摘要：On-policy Verbal Distillation用轨迹级蒸馏替代token级匹配蒸馏，使用离散的教师评分，降低内存占用并在网络搜索问答和数学推理上提升学生模型性能。 ### From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation - Venue: ICLR 2026 - Authors: Yuxin Jiang, Yufei Wang, Qiyuan Zhang, Xingshan Zeng, Liangyou Li, Jierun Chen, Chaofan Tao, Haoli Bai, Lifeng Shang - Paper: https://arxiv.org/abs/2601.18533 - Code: https://github.com/YJiangcm/RLVRR RLVRR turns single verifiable end rewards into ordered reference-derived reward chains (content + style) to make RL for open-ended generation more efficient and less prone to reward hacking. 中文摘要：RLVRR将传统RL里单一的可验证终端奖励转化为有序的参考衍生奖励链（内容+风格），使开放式生成的强化学习更高效且更不易出现奖励破解。 ### PhyX: Does Your Model Have the "Wits" for Physical Reasoning? - Venue: TMLR 2026 - Authors: Hui Shen, Taiqiang Wu, Qi Han, Yunta Hsieh, Jizhou Wang, Yuyue Zhang, Yuxin Cheng, Zijian Hao, Yuansheng Ni, Xin Wang, Zhongwei Wan, Kai Zhang, Wendong Xu, Jing Xiong, Ping Luo, Wenhu Chen, Chaofan Tao, Zhuoqing Mao, Ngai Wong - Paper: https://arxiv.org/abs/2505.15929 - Code: https://github.com/killthefullmoon/PhyX PhyX introduces a 3K-question multimodal benchmark spanning six physics domains to test physics-grounded reasoning, revealing large gaps between top models and human experts. 中文摘要：PhyX引入一个包含3000个问题的多模态基准，覆盖六个物理领域，用于测试基于物理的推理，观察到顶级模型与人类专家之间的较大差距。 ### Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies - Venue: NeurIPS 2024 - Authors: Chaofan Tao, Qian Liu, Longxu Dou, Niklas Muennighoff, Zhongwei Wan, Ping Luo, Min Lin, Ngai Wong - Paper: https://arxiv.org/abs/2407.13623 - Code: https://github.com/sail-sg/scaling-with-vocab The paper derives compute-optimal vocabulary scaling laws and shows larger models benefit from larger vocabularies, improving downstream accuracy at fixed FLOPs. 中文摘要：推导了计算最优的词表规模扩展规律，证明更大模型需要更大词表，在相同计算量下提升下游任务准确率。 ### RoboCodeX: Multimodal Code Generation for Robotic Behavior Synthesis - Venue: ICML 2024 - Authors: Yao Mu, Junting Chen, Qinglong Zhang, Shoufa Chen, Qiaojun Yu, Chongjian Ge, Runjian Chen, Zhixuan Liang, Mengkang Hu, Chaofan Tao, Peize Sun, Haibao Yu, Chao Yang, Wenqi Shao, Wenhai Wang, Jifeng Dai, Yu Qiao, Mingyu Ding, Ping Luo - Paper: https://arxiv.org/abs/2402.16117 RoboCodeX is a tree-structured multimodal code generation framework that decomposes high-level instructions into object-centric manipulation units with affordance and safety constraints, generalizing across robot platforms and reaching state-of-the-art results on manipulation and navigation tasks in simulation and on real robots. 中文摘要：RoboCodeX提出树状结构的多模态代码生成框架，将高层指令分解为带有可供性与安全约束的以物体为中心的操作单元，借助代码生成实现跨机器人平台泛化，在仿真与真实机器人的操作和导航任务上取得领先结果。 ### Autoregressive Models in Vision: A Survey - Venue: TMLR 2025 - Authors: Jing Xiong, Gongye Liu, Lun Huang, Chengyue Wu, Taiqiang Wu, Yao Mu, Yuan Yao, Hui Shen, Zhongwei Wan, Jinfa Huang, Chaofan Tao, Shen Yan, Huaxiu Yao, Lingpeng Kong, Hongxia Yang, Mi Zhang, Guillermo Sapiro, Jiebo Luo, Ping Luo, Ngai Wong - Paper: https://arxiv.org/abs/2411.05902 - Repository: https://github.com/ChaofanTao/Autoregressive-Models-in-Vision-Survey A comprehensive survey of visual autoregressive modeling, covering representation strategies and applications across image, video, 3D, and multimodal generation, plus open challenges. 中文摘要：全面综述视觉自回归模型的建模方法，涵盖表示策略及其在图像、视频、3D和多模态生成中的应用，并讨论开放挑战。 ### SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning - Venue: NeurIPS 2025 - Authors: Zhongwei Wan, Zhihao Dou, Che Liu, Yu Zhang, Dongfei Cui, Qinjian Zhao, Hui Shen, Jing Xiong, Yi Xin, Yifan Jiang, Chaofan Tao, Yangfan He, Mi Zhang, Shen Yan - Paper: https://arxiv.org/abs/2506.01713 - Code: https://github.com/SUSTechBruce/SRPO_MLLMs SRPO is a reflection-aware RL framework that teaches MLLMs self-reflection and then optimizes concise, meaningful reflections with GRPO-style rewards to improve multimodal reasoning. 中文摘要：SRPO是一个反思感知的多模态RL框架，先训练多模态大模型学会自我反思，再用GRPO式奖励优化简洁有效的反思，从而提升多模态推理能力。 ### The Synergy Dilemma of Long-CoT SFT and RL: Investigating Post-training Techniques for Reasoning VLMs - Venue: TMLR 2025 - Authors: Jierun Chen, Tiezheng Yu, Haoli Bai, Lewei Yao, Jiannan Wu, Kaican Li, Fei Mi, Chaofan Tao, Lei Zhu, Manyi Zhang, Xiaohui Li, Lu Hou, Lifeng Shang, Qun Liu - Paper: https://arxiv.org/abs/2507.07562 - Repository: https://github.com/JierunChen/SFT-RL-SynergyDilemma The study compares long-CoT SFT vs. RL for reasoning VLMs and finds their combination often fails to yield additive gains, instead trading off accuracy, verbosity, and style. 中文摘要：比较长链思维监督微调（SFT）与强化学习（RL）在视觉语言模型推理中的作用，发现两者组合往往无法带来叠加收益，而是在准确率、冗长程度和风格之间权衡。 ### NAVERO: Unlocking Fine-Grained Semantics for Video-Language Compositionality - Venue: arXiv 2024 - Authors: Chaofan Tao, Gukyeong Kwon, Varad Gunjal, Hao Yang, Zhaowei Cai, Yonatan Dukler, Ashwin Swaminathan, R Manmatha, Colin Jon Taylor, Stefano Soatto - Paper: https://arxiv.org/abs/2408.09511 NAVERO introduces the AARO benchmark and negative-text augmentation with a negative-augmented matching loss to improve video-language compositional understanding and retrieval. 中文摘要：NAVERO提出AARO基准，并通过负文本增强和负增强匹配损失，提升视频-语言模型对组合语义的理解和检索能力。 ## Efficient ML & AI Infrastructure ### Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models - Venue: EMNLP 2026 - Authors: Jing Xiong, Qi Han, Shansan Gong, Yunta Hsieh, Boyuan Zheng, Chengyue Wu, Chaofan Tao, Chenyang Zhao, Ngai Wong - Paper: https://arxiv.org/abs/2606.10537 - Code: https://github.com/menik1126/Prefilling-dLLM Prefilling-dLLM separates prefilling from decoding without training: it caches prefix KV representations by chunk, selects query-relevant chunks and tokens, and reuses the sparse cache to accelerate long-context diffusion language model inference. 中文摘要：Prefilling-dLLM无需额外训练，通过分块预填充缓存前缀KV表示，再选择与查询相关的文本块及块内关键token，并在去噪解码中复用稀疏缓存，从而降低扩散语言模型长上下文推理中的重复计算。 ### ATTS: Asynchronous Test-Time Scaling via Conformal Prediction - Venue: ICLR 2026 - Authors: Jing Xiong, Qiujiang Chen, Fanghua Ye, Zhongwei Wan, Chuanyang Zheng, Chenyang Zhao, Hui Shen, Alexander Hanbo Li, Chaofan Tao, Haochen Tan, Haoli Bai, Lifeng Shang, Lingpeng Kong, Ngai Wong - Paper: https://arxiv.org/abs/2509.15148 - Code: https://github.com/menik1126/Asynchronous-Test-Time-Scaling ATTS is an asynchronous, conformal-prediction-based test-time scaling method that cuts synchronization overhead and accelerates scaling with statistical guarantees and no accuracy loss. 中文摘要：ATTS是一种基于一致性预测的异步测试时扩展方法，减少同步开销，在尽可能最小化准确率损失的情况下加速扩展。 ### UNCOMP: Can Matrix Entropy Uncover Sparsity? - A Compressor Design from an Uncertainty-Aware Perspective - Venue: EMNLP 2025 - Authors: Jing Xiong, Jianghan Shen, Fanghua Ye, Chaofan Tao, Zhongwei Wan, Jianqiao Lu, Xun Wu, Chuanyang Zheng, Zhijiang Guo, Min Yang, Lingpeng Kong, Ngai Wong - Paper: https://doi.org/10.48448/v1xq-f328 - Code: https://github.com/menik1126/UNComp UNComp uses uncertainty (via truncated matrix entropy) to uncover KV-cache sparsity and perform adaptive compression for long-context inference, greatly shrinking cache while boosting throughput. 中文摘要：UNComp使用不确定性（通过截断矩阵熵）来发现KV缓存的稀疏性，并对长上下文推理进行自适应压缩，大幅缩小缓存同时提升吞吐量。 ### Find Your Optimal Teacher: Personalized Data Synthesis via Router-Guided Multi-Teacher Distillation - Venue: ACL (Main Conference) 2026 - Authors: Hengyuan Zhang, Shiping Yang, Xiao Liang, Chenming Shang, Yuxuan Jiang, Chaofan Tao, Jing Xiong, Hayden Kwok-Hay So, Ruobing Xie, Angel X Chang, Ngai Wong - Paper: https://arxiv.org/abs/2510.10925 - Code: https://github.com/rattlesnakey/PerSyn PerSyn "routes then generates" by assigning each prompt to its best teacher via a router that considers student learnability and teacher quality, improving synthetic-data distillation efficiency. 中文摘要：PerSyn采用"先路由后生成"策略，通过考虑学生学习能力和教师质量的路由器为每个提示分配最佳教师，提高合成数据蒸馏效率。 ### AnchorTP: Resilient LLM Inference with State-Preserving Elastic Tensor Parallelism - Venue: DATE 2026 - Authors: Wendong Xu, Chujie Chen, He Xiao, Kuan Li, Jing Xiong, Chen Zhang, Wenyong Zhou, Chaofan Tao, Yang Bai, Bei Yu, Ngai Wong - Paper: https://arxiv.org/abs/2511.11617 AnchorTP provides state-preserving elastic tensor parallelism with a GPU-resident daemon and bandwidth-aware migration planning, enabling fast recovery from single-GPU failures. 中文摘要：AnchorTP提供状态保持的弹性张量并行，通过GPU常驻守护进程和带宽感知的迁移规划，实现单GPU故障场景的快速恢复。 ### Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models - Venue: ICLR 2024 - Authors: Zhongwei Wan, Xinjian Wu, Yu Zhang, Yi Xin, Chaofan Tao, Zhihong Zhu, Xin Wang, Siqi Luo, Jing Xiong, Longyue Wang, Mi Zhang - Paper: https://arxiv.org/abs/2406.13035 - Code: https://github.com/AIoT-MLSys-Lab/D2O D2O dynamically compresses KV cache at layer and token levels with allocation and recall/merge compensation, saving memory and improving long-context throughput without fine-tuning. 中文摘要：D2O在层和token级别动态压缩KV缓存，通过分配和召回/合并补偿，无需微调即可节省内存并提升长上下文吞吐量。 ### CrossGET: Cross-Guided Ensemble of Tokens for Accelerating Vision-Language Transformers - Venue: ICML 2024 - Authors: Dachuan Shi, Chaofan Tao, Anyi Rao, Zhendong Yang, Chun Yuan, Jiaqi Wang - Paper: https://arxiv.org/abs/2305.17455 - Code: https://github.com/sdc17/CrossGET CrossGET accelerates vision-language Transformers by cross-modally matching and ensembling tokens with efficient soft matching, reducing compute with minimal accuracy loss. 中文摘要：CrossGET通过跨模态匹配和集成token，使用相似token匹配和合并加速视觉-语言Transformer，在最小准确率损失下减少计算量。 ### Rethinking Kullback-Leibler Divergence in Knowledge Distillation for Large Language Models - Venue: COLING 2024 - Authors: Taiqiang Wu, Chaofan Tao, Jiahao Wang, Runming Yang, Zhe Zhao, Ngai Wong - Paper: https://arxiv.org/abs/2404.02657 - Code: https://github.com/wutaiqiang/LLM_KD_AKL It shows forward and reverse KL behave similarly in LLM distillation with enough training but emphasize head vs. tail early; Adaptive KL combines them to improve generation quality and diversity. 中文摘要：研究表明在充分训练下，前向和反向KL在大语言模型蒸馏中表现相似，但早期分别关注分布头部和尾部；自适应KL结合两种蒸馏方式的优势，以提升模型的生成质量和多样性。 ### Structured Pruning for Efficient Generative Pre-trained Language Models - Venue: Findings of ACL 2023 - Authors: Chaofan Tao, Lu Hou, Haoli Bai, Jiansheng Wei, Xin Jiang, Qun Liu, Ping Luo, Ngai Wong - Paper: https://doi.org/10.18653/v1/2023.findings-acl.692 It proposes structured pruning across multiple generative-PLM components using learnable masks and sparse training to extract size-flexible subnets with significant speedups. 中文摘要：提出对生成式预训练语言模型多个组件的结构化剪枝，使用可学习掩码和稀疏训练提取尺寸灵活的子网络，实现显著加速。 ### Compression of Generative Pre-trained Language Models via Quantization - Venue: ACL 2022 - Award: ACL 2022 Outstanding Paper Award - Authors: Chaofan Tao, Lu Hou, Wei Zhang, Lifeng Shang, Xin Jiang, Qun Liu, Ping Luo, Ngai Wong - Paper: https://arxiv.org/abs/2203.10705 It adapts quantization to generative PLMs （e.g. GPT-2, BART）, addressing embedding homogenization with token-level contrastive distillation and module-wise dynamic scaling to achieve 13–14× compression at near full precision. 中文摘要：将量化方法适配到生成式预训练语言模型（e.g. GPT-2, BART），通过token级对比蒸馏解决嵌入同质化问题，并用模块级动态缩放实现接近全精度的13-14倍压缩。 ### UPop: Unified and Progressive Pruning for Compressing Vision-Language Transformers - Venue: ICML 2023 - Authors: Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, Jiaqi Wang - Paper: https://arxiv.org/abs/2301.13741 - Code: https://github.com/sdc17/UPop UPop is a unified progressive pruning framework that searches multimodal subnets in a continuous space and progressively retrains to achieve high compression for vision-language Transformers. 中文摘要：UPop是一个统一的渐进式剪枝框架，在连续空间中搜索多模态子网络并逐步重训练，实现视觉-语言Transformer的高压缩率。