Mufei Li (李牧非)'s Homepage

I’m a fourth-year PhD student in machine learning at Georgia Institute of Technology, advised by Prof. Pan Li. I’m broadly interested in algorithmic ideas and algorithm-system co-design for large language models and agents, including but not limited to long-context reasoning, KV cache manipulation, reinforcement learning, on-policy distillation (OPD), etc.

Prior to that, I was a software development engineer (SDE) at Amazon Web Services (AWS) Shanghai AI Lab. I received my bachelor’s degree in Honors Math from NYU Shanghai.

I am always open to chat about exciting internship and full-time opportunities. Please feel free to reach out if you have any openings!

News

Nov 12, 2026	Check our new paper on multi-turn agentic on-policy distillation, Which Successes Make Good References? Action-Sequence Alignment for Multi-Turn On-Policy Distillation
Sep 24, 2026	Parallel-Synthesis has been accepted by NeurIPS 2026! Congratulations to Shikun, and all the co-authors!
Jul 8, 2026	Struc-EMB has been accepted by COLM 2026! Congratulations to Shikun, and all the co-authors!
May 21, 2026	Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows has been accepted by ICML 2026 workshop on Graph Foundation Models: A New Era for Graph Machine Learning! Congratulations to Shikun, and all the co-authors!
May 18, 2026	I’m starting a new position as Applied Scientist Intern at AWS! I’m based in Seattle, Washington.

Selected Publications

arXiv

Which Successes Make Good References? Action-Sequence Alignment for Multi-Turn On-Policy Distillation

Mufei Li, Nikki Lijing Kuang, Zhuoer Wang, Samson Koelle, Jie Yuan, Manqing Mao, Hong Wang, Yanjun Lin, James Feng, Daniel Edmiston, Pan Li, and Wei Niu

arXiv preprint arXiv:2611.13135, 2026

arXiv Bib

@article{li2026successes,
  title = {Which Successes Make Good References? Action-Sequence Alignment for Multi-Turn On-Policy Distillation},
  author = {Li, Mufei and Kuang, Nikki Lijing and Wang, Zhuoer and Koelle, Samson and Yuan, Jie and Mao, Manqing and Wang, Hong and Lin, Yanjun and Feng, James and Edmiston, Daniel and Li, Pan and Niu, Wei},
  journal = {arXiv preprint arXiv:2611.13135},
  year = {2026},
}

NeurIPS

Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows

Shikun Liu, Mufei Li, Dongqi Fu, Haoyu Wang, Yinglong Xia, Hong Li, Hong Yan, and Pan Li

2026

arXiv Bib Code

@article{liu2026towards,
  title = {Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows},
  author = {Liu, Shikun and Li, Mufei and Fu, Dongqi and Wang, Haoyu and Xia, Yinglong and Li, Hong and Yan, Hong and Li, Pan},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026},
}

ICML MemFM Oral

KVEraser: Learning to Steer KV Cache for Efficient Localized Context Erasing

Mufei Li, Shikun Liu, Dongqi Fu, Haoyu Wang, Yinglong Xia, Hong Li, Hong Yan, and Pan Li

arXiv preprint arXiv:2606.17034, 2026

arXiv Bib Code

@article{li2026kveraser,
  title = {KVEraser: Learning to Steer KV Cache for Efficient Localized Context Erasing},
  author = {Li, Mufei and Liu, Shikun and Fu, Dongqi and Wang, Haoyu and Xia, Yinglong and Li, Hong and Yan, Hong and Li, Pan},
  journal = {arXiv preprint arXiv:2606.17034},
  year = {2026},
}

ICLR Oral

Reducing Belief Deviation in Reinforcement Learning for Active Reasoning of LLM Agents

Deyu Zou, Yongqiang Chen, Jianxiang Wang, Garry YANG, Mufei Li, Qing Da, James Cheng, Pan Li, and Yu Gong

In The Fourteenth International Conference on Learning Representations, 2026

arXiv Bib Code

@inproceedings{zoureducing,
  title = {Reducing Belief Deviation in Reinforcement Learning for Active Reasoning of LLM Agents},
  author = {Zou, Deyu and Chen, Yongqiang and Wang, Jianxiang and YANG, Garry and Li, Mufei and Da, Qing and Cheng, James and Li, Pan and Gong, Yu},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year = {2026},
}

ICLR
Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation

Mufei Li, Siqi Miao, and Pan Li

In International Conference on Learning Representations, 2025

Mufei Li and Siqi Miao contributed equally to this work. This work was selected for oral presentation at [ICLR 2025 workshop on Foundation Models in the Wild](https://fm-wild-community.github.io/)

arXiv Bib Code ...⭐
@inproceedings{li2025subgraphrag, title = {Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation}, author = {Li, Mufei and Miao, Siqi and Li, Pan}, booktitle = {International Conference on Learning Representations}, year = {2025}, github_stars = {Graph-COM/SubgraphRAG}, }

ICLR Spotlight

LayerDAG: A Layerwise Autoregressive Diffusion Model for Directed Acyclic Graph Generation

Mufei Li, Viraj Shitole, Eli Chien, Changhai Man, Zhaodong Wang, Srinivas Sridharan, Ying Zhang, Tushar Krishna, and Pan Li

In International Conference on Learning Representations, 2025

Selected for spotlight presentation at the ICLR 2025 workshop ["Will Synthetic Data Finally Solve the Data Access Problem?"](https://synthetic-data-iclr.github.io/)

arXiv Bib Code ...⭐

@inproceedings{li2024layerdag,
  title = {Layer{DAG}: A Layerwise Autoregressive Diffusion Model for Directed Acyclic Graph Generation},
  author = {Li, Mufei and Shitole, Viraj and Chien, Eli and Man, Changhai and Wang, Zhaodong and Sridharan, Srinivas and Zhang, Ying and Krishna, Tushar and Li, Pan},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  note = {Selected for spotlight presentation at the ICLR 2025 workshop ["Will Synthetic Data Finally Solve the Data Access Problem?"](https://synthetic-data-iclr.github.io/)},
  github_stars = {Graph-COM/LayerDAG},
}