Ph.D. Student

Computer Science Department

University of Virginia

Email: [email protected] or [email protected]

Now I’m working on:

• LLM Reasoning
• RL for LLMs

X(Twitter) | Linkedin | CV

Github | Google Scholar

Hi!👋 I’m Xinyu (/ʃɪn.ju:/) Zhu, ****a second-year Ph.D. student at the University of Virginia, advised by Prof. ****Yu Meng. I am broadly interested in LLMs and ML, especially in improving LLMs via reinforcement learning. My long-term research goal is to enable human expert level reasoning, decision-making and cognitive intelligence for neural models and systems. I received my master degree at Tsinghua University, advised by Prof. ****Yujiu Yang.

✨News

• [09/2025] Two works on how negative-reward-only RL can improve LLM reasoning and transferring small RL-tuned models’ reasoning capability to large base model are accepted to NeurIPS 2025! ****🎉 See you in San Diego! 🤩
• [06/2025] Check out our new preprint and X thread on how negative-reward-only RL can improve LLM reasoning! 🚀
• [06/2025] Our AdaDecode is accepted at ICML 2025! 🎉
• [05/2025] I moved to Cupertino and started my internship at Apple AIML, looking forward to meeting new friends in the Bay Area! 🚀
• [03/2025] One survey on the honesty of LLMs is accepted at TMLR! 🎉
• [01/2025] One paper on evaluating LMM's reasoning via chart-to-code generation ****is accepted at ICLR 2025! 🎉
• [10/2024] Check out our comprehensive survey on the honesty of LLMs! 🎉 Also checkout our accompanied github repo! PRs welcome!
• [09/2024] One paper on improving MoE models' reasoning via self-contrast has been accepted at NeurIPS 2024, congrats to all the co-authors🎉! See you in Vancouver!
• Old news

👨‍🎓Education

• University of Virginia 2024 -- Present Ph.D. in Computer Science
• Tsinghua University 2021 -- 2024 Master in Electronic and Information Engineering
• Xidian University 2017 -- 2021 Bachelor in Electronic Science and Technology

📝Selected Works (* indicates equal contribution)

The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning

NeurIPS 2025

Xinyu Zhu, Mengzhou Xia, Zhepei Wei, Wei-Lin Chen, Danqi Chen, Yu Meng

[paper][code][X thread]