Jiayan Fu
School of Mathematical Sciences, Peking University
I am a master’s student in Big Data at the School of Mathematical Sciences, Peking University, advised by Dongyan Zhao. Before joining PKU, I earned a bachelor’s degree in Mathematics.
My research interests include Large Language Models and Reinforcement Learning, particularly understanding the mathematical mechanisms behind reinforcement learning for large language models through probability theory. Before shifting my focus to AI, my main interests in mathematics were analysis and probability theory.
My recent work mainly focuses on credit assignment in long-horizon tasks.
I am always happy to discuss research ideas in reinforcement learning for LLMs. You can reach me at fujiayan@live.com.
Selected Papers
- PACT: From Credit Assignment to Critic AlignmentTechnical Report , Sep 2026
We prove that three regularity conditions—Completeness, Prefix Consistency, and Neutrality—uniquely determine token-level credit, providing a unified perspective on OPD, RLOO, and GAE. We establish approximate credit sparsity under bounded outcome rewards and show how intermediate critic errors can become comparable to the underlying credit. These findings motivate Policy Aligned Critic Training (PACT), which uses an Actor-then-Critic update order and importance sampling correction to better align the critic with the updated policy. PACT achieves 72.87% average accuracy across four agentic mathematical reasoning benchmarks and 67.4% on SWE-bench Verified, outperforming GRPO and PPO in both settings.
Selected Blogs
Are SAO’s Critic Updates Really Better? — October 9, 2026
Why Normalize GRPO Advantages? A Variance-Stabilizing Perspective — August 3, 2026
Token-Mean Aggregation Is Asymptotically Unbiased — May 27, 2026
Experience
Foundation Model Algorithm Intern · RedNote, AllSpark April 2026 – Present
My work mainly focuses on agentic reinforcement learning, especially its algorithms.