🌈 I am Haolei Xu (徐皓雷), a second-year PhD student at Zhejiang University, College of Computer Science and Technology. Previously, I received my Bachelor’s degree from the School of Computer Science, Harbin Institute of Technology.

📌 My research focuses on LLM Reasoning, Reinforcement Learning, and Interpretability.

📢 Feel free to reach out via email: xuhaolei@zju.edu.cn.

🔥 News

📝 Publications

Full publications are on my Google Scholar profile. *: Equal contribution. †: Project leader. ‡: Corresponding author.

ACL 2026

🔍 Identifies the Seeing but Not Thinking phenomenon in multimodal MoE models and proposes a routing-guided fix.

Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts

Haolei Xu^*, Haiwen Hong^*,†, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen^‡, Weiming Lu^‡, Yueting Zhuang

• Discovers that multimodal MoE models correctly perceive visual content yet fail at reasoning: 68–73% of failures stem from reasoning errors, not perception.
• Reveals Routing Distraction: image inputs induce divergence in middle-layer routing, diverting computation away from domain reasoning experts.
• Proposes routing-guided intervention achieving up to +3.17% on complex visual reasoning across 3 MoE models and 6 benchmarks.

[paper] [PaperWeekly]

NeurIPS 2025

🧩 Addresses Thought Leap in CoT datasets by automatically detecting and bridging missing intermediate reasoning steps.

Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning

Haolei Xu^*, Yuchen Yan^*, Yongliang Shen^‡, Wenqi Zhang, Guiyang Hou, Shengpei Jiang, Kaitao Song, Weiming Lu^‡, Jun Xiao, Yueting Zhuang

• Identifies Thought Leap — omitted intermediate steps in CoT chains — causing up to 27.83% lower performance ceilings on models trained with such gaps.
• Constructs ScaleQM+ and trains CoT-Bridge to automatically detect and fill missing reasoning steps, restoring completeness and coherence of CoT data.
• Fine-tuned models achieve up to +5.87% on NuminaMath; also improves distillation quality (+3.02%) and RL cold-start initialization (+3.1%).

[paper] [code] [page] [机器之心]

arXiv

🚀 A unified LLM steering framework built on vLLM with 10.8–22.3× speedup over existing methods.

EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering

Haolei Xu, Xinyu Mei, Yuchen Yan, Rui Zhou, Wenqi Zhang, Weiming Lu^‡, Yueting Zhuang, Yongliang Shen^‡

• Modular, pluggable architecture addressing inefficiency and limited extensibility of prior steering frameworks.
• 10.8–22.3× inference speedup via deep vLLM integration; 81–91% baseline throughput under multi-vector use.
• Overthinking mitigation (tokens ↓40%), hallucination reduction (+12% accuracy), 8 pre-built domains.

[paper] [code] [机器之心]

Conference Papers

Haolei Xu^*, Haiwen Hong^*,†, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen^‡, Weiming Lu^‡, Yueting Zhuang, "Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts". In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2026. [paper] [PaperWeekly]

Haolei Xu^*, Yuchen Yan^*, Yongliang Shen^‡, Wenqi Zhang, Guiyang Hou, Shengpei Jiang, Kaitao Song, Weiming Lu^‡, Jun Xiao, Yueting Zhuang, "Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning". In Advances in Neural Information Processing Systems (NeurIPS), 2025. [paper] [code] [page] [机器之心]

Haoran Zhao^*, Yuchen Yan^*, Yongliang Shen^‡, Haolei Xu, Wenqi Zhang, Kaitao Song, Jian Shao, Weiming Lu, Jun Xiao, Yueting Zhuang, "Let LRMs Break Free from Overthinking via Self-Braking Tuning". In Advances in Neural Information Processing Systems (NeurIPS), 2025. [paper] [code] [page] [量子位]

Siqi Chen^*, Xinyu Dong^*, Haolei Xu, Xingyu Wu, Fei Tang, Hang Zhang, Yuchen Yan, Linjuan Wu, Wenqi Zhang, Guiyang Hou, Yongliang Shen, Weiming Lu, Yueting Zhuang, "SVGenius: Benchmarking LLMs in SVG Understanding, Editing and Generation". In Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM), 2025. [paper] [code] [page]

Haoyuan Ma, Yongliang Shen^‡, Hengwei Liu, Wenqi Zhang, Haolei Xu, Qiuying Peng, Jun Wang, Weiming Lu^‡, "DB-Explore: Automated Database Exploration and Instruction Synthesis for Text-to-SQL". In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025. [paper]

Preprints & Under Submission

Fei Tang^*, Haolei Xu^*, Hang Zhang^*, Siqi Chen^*, Xingyu Wu^*, Yongliang Shen^‡, Wenqi Zhang, Guiyang Hou, Zeqi Tan, Yuchen Yan, Kaitao Song, Jian Shao, Weiming Lu, Jun Xiao, Yueting Zhuang, "A Survey on (M)LLM-based GUI Agents". arXiv preprint arXiv:2504.13865, 2025. [paper] [code]

Haolei Xu, Xinyu Mei, Yuchen Yan, Rui Zhou, Wenqi Zhang, Weiming Lu^‡, Yueting Zhuang, Yongliang Shen^‡, "EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering". arXiv preprint arXiv:2509.25175, 2025. [paper] [code] [机器之心]

🎓 Education

• Ph.D. in Computer Science — Zhejiang University
  • Time: Sep 2024 – Present.
  • College of Computer Science and Technology.
• B.S. in Computer Science — Harbin Institute of Technology
  • Time: Sep 2020 – Jun 2024.
  • School of Computer Science.