Yi Hu (胡逸)

yihu.jpg

Email: huyi2002 (at) stu (dot) pku (dot) edu (dot) cn

I am now a third-year Ph.D. student at the Institute for Artificial Intelligence, Peking University, advised by Prof. Muhan Zhang. Welcome to visit our group page: Graphpku!

I received my B.S. in Physics from School of Physics, Peking University, where I was fortunate to work with Prof. Huichao Song on ML applications in heavy-ion collision.

I am dedicated to exploring the reasoning mechanisms of large language models (LLMs) and researching how to enhance the models’ reasoning capabilities to the level of human experts. Additionally, I have a keen interest in various topics related to LLMs, including efficiency, alignment, and applications across different downstream domains etc. If you share these research interests, please feel free to get in touch!

Explore my group meeting slides and research notes.

News

Oct 14, 2026 🔥🔥 When context conflicts with what an LLM already knows, which should it trust? 🤔 In Which Knowledge Should LLMs Trust?, we uncover a transferable steering direction that helps control this choice.
Oct 11, 2026 🔥🔥 In continual learning, can LLMs combine new knowledge with what they already know? Our new paper Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and Composition explores conflicts and composition across the base model, adapters, and context in continual learning.
Sep 30, 2026 🔥🔥 Do shared weights mean the same computation at every loop? 🔁 In our new paper: Shared Weights, Selected Computations, we study how hidden states and attention route different computations through looped Transformers.
Apr 07, 2026 Our survey on reasoning model mechanisms (Oral) and SubTokenTest were accepted by ACL 2026 Main!
Jan 29, 2026 Two preprints: SubTokenTest (benchmarking real-world sub-token understanding for LLMs) and the first Reasoning Model Mechanisms Survey.
Sep 23, 2025 🥳🥳 Two papers accepted at NeurIPS 2025, check out here: Meta-RFFT and PHYBench. One paper accepted at NeurIPS Mech Interp Workshop: What Affects the Effective Depth of Large Language Models?
Apr 23, 2025 🔥🔥 We collaborate with School of Physics, Peking University and release PHYBench, a physical reasoning benchmark for modern LLMs. Covering mechanics, electromagnetism, thermodynamics, optics, modern physics, and advanced physics, the benchmark spans difficulty levels from high school exercises to undergraduate problems and Physics Olympiad challenges. Check out our paper: PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models.
Feb 17, 2025 How to enhance the rule-following ability of LLMs? 🤔 We propose Meta Rule-Following Fine-Tuning (Meta-RFFT) to improve the cross-task transferability of rule-following abilities. We construct a dataset of 88 length generalization tasks and show Meta-RFFT helps models to outperform baselines in both downstream fine-tuning and few-shot prompting scenarios. 👋 Check out our new paper: Training Large Language Models to be Better Rule Followers!
Jan 23, 2025 NUPA got accepted by ICLR 2025🥳🥳!! Check out our paper: Number Cookbook: Number Understanding of Language Models and How to Improve It!
Nov 07, 2024 Why 9.9 > 9.11 is so hard for LLMs? 😫 Check out our new paper: Number Cookbook: Number Understanding of Language Models and How to Improve It, where we introduce a comprehensive benchmark covering four common numerical representations and 17 distinct numerical tasks and investigate the numerical understanding and processing ability (NUPA) of LLMs.

Selected publications

  1. preprint
    Which Knowledge Should LLMs Trust? Uncovering a Generalizable Steering Direction Across Knowledge Conflicts
    Yibin Lou, Yi Hu, Jianhui Chen, Chenpeng Wang, Muhan Zhang, and Liangming Pan
    arXiv preprint arXiv:2610.20067, 2026
    Paper
  2. preprint
    Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and Composition
    Jiaqi Zheng, Yi Hu, Yadong Mu, and Muhan Zhang
    arXiv preprint arXiv:2610.15677, 2026
    Paper
  3. preprint
    Shared Weights, Selected Computations: How Looped Transformers Route What Each Loop Does
    Jiaju Wu, Yi Hu, and Muhan Zhang
    arXiv preprint arXiv:2609.39892, 2026
    Paper Code
  4. ACL 2026 Oral
    Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures
    Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, and Liangming Pan
    In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, 2026
    Paper Code
  5. NeurIPS WS
    What Affects the Effective Depth of Large Language Models?
    Yi Hu, Cai Zhou, and Muhan Zhang
    2025
    Paper
  6. NIPS2025
    PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models
    Shi Qiu, Shaoyang Guo, Zhuo-Yang Song, Yunbo Sun, Zeyu Cai, Jiashen Wei, Tianyu Luo, Yixuan Yin, Haoxu Zhang, Yi Hu, Chenyang Wang, Chencheng Tang, Haoling Chang, Qi Liu, Ziheng Zhou, Tianyu Zhang, Jingtian Zhang, Zhangyi Liu, Minghao Li, Yuku Zhang, Boxuan Jing, Xianqi Yin, Yutong Ren, Zizhuo Fu, Weike Wang, Xudong Tian, Anqi Lv, Laifu Man, Jianxiang Li, Feiyu Tao, Qihua Sun, Zhou Liang, Yushu Mu, Zhongxuan Li, Jing-Jun Zhang, Shutao Zhang, Xiaotian Li, Xingqi Xia, Jiawei Lin, Zheyu Shen, Jiahang Chen, Qiuhao Xiong, Binran Wang, Fengyuan Wang, Ziyang Ni, Bohan Zhang, Fan Cui, Changkun Shao, Qing-Hong Cao, Ming-xing Luo, Muhan Zhang, and Hua Xing Zhu
    2025
    Paper
  7. NIPS2025
    Beyond Single-Task: Robust Multi-Task Length Generalization for LLMs
    Yi Hu, Shijia Kang, Haotong Yang, Haotian Xu, and Muhan Zhang
    2025
    Paper
  8. ICLR2025
    Number Cookbook: Number Understanding of Language Models and How to Improve It
    Haotong Yang, Yi Hu, Shijia Kang, Zhouchen Lin, and Muhan Zhang
    2025
    Paper Code
  9. ICML2024
    Case-Based or Rule-Based: How Do Transformers Do the Math?
    Yi Hu, Xiaojuan Tang, Haotong Yang, and Muhan Zhang
    2024
    Paper Code Poster