I’m a 5th-year CS PhD student at Cornell University, advised by Prof. Kilian Weinberger and Wen Sun. My research focuses on LLMs for math reasoning and theorem proving, efficient LLM evaluation, and RLHF for personalization and mitigating hallucinations. I graduated from the University of Toronto’s Engineering Science program (Machine Intelligence option), where I completed my thesis with Prof. Roger Grosse and was fortunate to have worked with Prof. Nicolas Papernot and Scott Sanner.

During my PhD, I’ve had a lot of fun interning at Meta (MSL), Google DeepMind (with Seb Arnold and Fei Sha), and Google Research (with Yuhuai Wu and Christian Szegedy). I am supported by an NSERC CGS-D Fellowship, Cornell University Fellowship and NSF NAIRR Pilot Award.

Selected Papers

Q♯: Provably Optimal Distributional RL for LLM Post-Training

Jin Peng Zhou*, Kaiwen Wang*, Jonathan Chang*, Zhaolin Gao, Nathan Kallus, Kilian Weinberger, Kianté Brantley, Wen Sun

NeurIPS 2025

Provably optimal distributional RL method for KL-regularized RL, which we apply to post-train language models.

PDF Code

Value-Guided Search for Efficient Chain-of-Thought Reasoning

Kaiwen Wang*, Jin Peng Zhou*, Jonathan Chang*, Zhaolin Gao, Nathan Kallus, Kianté Brantley, Wen Sun

NeurIPS 2025

Block-wise value-guided search for scaling test-time compute of reasoning models (DeepSeek-R1 series).

PDF Code

Graders Should Cheat: Privileged Information Enables Expert-Level Automated Evaluations

Jin Peng Zhou, Sébastien M. R. Arnold, Nan Ding, Kilian Weinberger, Nan Hua, Fei Sha

EMNLP Main Conference 2025

Privileged information significantly improves LLM judge performance on diverse tasks.

PDF

On Speeding Up Language Model Evaluation

Jin Peng Zhou*, Christian Belardi*, Ruihan Wu*, Travis Zhang, Carla Gomes, Wen Sun, Kilian Weinberger

ICLR 2025

Adaptive evaluation methods such as multi-arm bandits cut LLM evaluation costs for identifying the best model.

PDF Code

Gemma 2: Improving Open Language Models at a Practical Size

Gemma Team

Arxiv 2025

Strong open-weight model released by Google. Contributed to the evaluation of model checkpoints.

PDF Checkpoints

Code Repair with LLMs Gives an Exploration-Exploitation Tradeoff

Hao Tang, Keya Hu, Jin Peng Zhou, Si Cheng Zhong, Wei-Long Zheng, Xujie Si, Kevin Ellis

NeurIPS 2024

Balancing exploration and exploitation in iterative code repair with a bandit-based refinement strategy.

PDF

Don't Trust: Verify - Grounding LLM Quantitative Reasoning with Autoformalization

Jin Peng Zhou, Charles Staats, Wenda Li, Christian Szegedy, Kilian Weinberger, Yuhuai Wu

ICLR 2024

Verifying LLM generated solutions with formal theorem prover enhances majority voting accuracy.

PDF Code

REFACTOR: Learning to Extract Theorems from Proofs

Jin Peng Zhou*, Yuhuai Wu*, Qiyang Li, Roger Grosse

ICLR 2024

Extracting and reusing theorems from Metamath proofs augments the library and improves prover accuracy.

PDF Code

Magnushammer: A Transformer-based Approach to Premise Selection

Maciej Mikuła*, Szymon Antoniak*, Szymon Tworkowski*, Albert Qiaochu Jiang, Jin Peng Zhou, Christian Szegedy, Łukasz Kuciński, Piotr Miłoś, Yuhuai Wu

ICLR 2024

Learning neural premise selection outperforms off-the-shelf automated theorem prover.

PDF Code

Unsupervised Out-of-Distribution Detection with Diffusion Inpainting

Zhenzhen Liu*, Jin Peng Zhou*, Yufan Wang, Kilian Q. Weinberger

ICML 2023

Detecting OOD samples by training a diffusion model on in-distribution data and using inpainting to expose anomalies.

PDF Code

Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs

Albert Q. Jiang*, Sean Welleck*, Jin Peng Zhou*, Wenda Li, Jiacheng Liu, Mateja Jamnik, Timothée Lacroix, Yuhuai Wu, Guillaume Lample

ICLR 2023

Oral

DSP maps natural language proofs into formal sketches that guide automated theorem provers, improving success rates on competition-level problems.

PDF Code

Does Label Differential Privacy Prevent Label Inference Attacks?

Ruihan Wu*, Jin Peng Zhou*, Kilian Q Weinberger, Chuan Guo

AISTATS 2023

Label differential privacy limits an adversary’s advantage in label inference attacks, rather than guaranteeing low absolute accuracy.

PDF Code

TAFA: Two-headed Attention Fused Autoencoder for Context-Aware Recommendations

Jin Peng Zhou*, Zhaoyue Cheng*, Felipe Perez, Maksims Volkovs

RecSys 2020

TAFA fuses reviews and implicit feedback to reduce popularity bias and improve personalized, interpretable recommendations.

PDF Code