Zhepei Wei (魏哲培)

Ph.D. Student Computer Science Department University of Virginia Email: zhepei.wei [AT] virginia [DOT] edu [Google Scholar] [GitHub]

About me

Hello! My name is Zhepei (/dʒɜ:peɪ/) Wei, and I also go by Bruce. I am a CS PhD student at the University of Virginia (UVA), advised by Prof. Yu Meng. Before joining UVA, I received my B.S. and M.S. degrees in Computer Science from Jilin University in 2019 and 2022. During my academic journey, I was fortunate to be advised by Prof. Hongning Wang and Prof. Yi Chang, and to gain industry research experience through internships at Amazon, Meta, and Microsoft Research, where I worked closely with Dr. Lihong Li, Dr. Luna Dong, Dr. Scott Yih, and Dr. Jianfeng Gao. I am also a recipient of the Capital One Ph.D. Fellowship, the UVA Copenhaver Charitable Trust Bicentennial Fellowship and the John A. Stankovic Outstanding Graduate Research Award.

For more information, please check my CV.

On the job market. I'm open to both academic and industry positions starting in Spring/Fall 2027. Please feel free to reach out if interested!

What's New

Research Interests

My research investigates the learning foundations of large language models (LLMs) and their applications to practical problems (e.g., logical reasoning, information seeking, computer use and coding), with a focus on efficiency, trustworthiness, and agentic capability.

Efficiency across data, inference and training:

• Fully self-synthetic data for alignment (Self-Alignment Optimization [Yin et al., ACL 2026])

• Adaptive layer parallelism for faster decoding (AdaDecode [Wei et al., ICML 2025])

• Effective reinforcement learning for reasoning with drastically reduced training steps (RELEX [Wei et al., arXiv 2026]) and with only negative trajectories (Negative RLVR [Zhu et al., NeurIPS 2025])

Trustworthiness through grounded generation, truthfulness-driven optimization and reliable evaluation:

• Retrieval-augmented generation with rationales (InstructRAG [Wei et al., ICLR 2025])

• Directly optimizing truthfulness rather than accuracy to reduce hallucinations (TruthRL [Wei et al., ICML 2026])

• Understanding self-preference of LLM evaluators ([Chen et al., arXiv 2026]).

Agentic capability (e.g., tool calling, planning) for real-world scenarios such as computer use, search and coding:

• End-to-end multi-turn reinforcement learning for computer use agents (WebAgent-R1 [Wei et al., EMNLP 2025])

• Training better search agents beyond outcome-only rewards (Decoupling Search-and-Answering [Wang et al., arXiv 2026])

• Iterative self-improvement for coding agents by co-evolving the model and harness ([Wei et al., work in progress])

I also worked on reinforcement learning (RL) theory, where I introduced incentivized collaboration for multi-agent decision-making and proposed efficient communication algorithms with rigorous mathematical guarantees (e.g., Inc-FedUCB [Wei et al., NeurIPS 2023], Truth-FedBan [Wei et al., ICLR 2024]).

Selected Awards & Honors

• 2026. Capital One Ph.D. Fellowship, Capital One

• 2026. Outstanding Reviewer Award, Association for Computational Linguistics (ACL)

• 2026. All-University Graduate Teaching Award, University of Virginia

• 2026. Silver Reviewer Award, International Conference on Machine Learning (ICML)

• 2025. Top Reviewer Award, International Conference on Machine Learning (ICML)

• 2025. Travel grant, International Conference on Learning Representations (ICLR)

• 2024. Copenhaver Charitable Trust Bicentennial Fellowship, University of Virginia

• 2024. John A. Stankovic Graduate Research Award, University of Virginia

• 2024. Travel grant, International Conference on Learning Representations (ICLR)

• 2023. NeurIPS Scholar Award, Neural Information Processing Systems (NeurIPS) Foundation

• 2022. Computer Science Scholar Fellowship, University of Virginia