Hi there! I am Yu Li (李煜, /juː/), a Senior Research Scientist at Salesforce AI Research. I completed my Ph.D. in Computer Science at Columbia University in 2025, advised by Prof. Zhou Yu.

My current research focuses on LLM agents and agent post-training, including efficient inference, memory-based learning, and reinforcement learning. I also study user simulation and long-horizon AI behavior.

Previously I worked on conversational AI, including dialogue systems, dialogue summarization, language model alignment, and the use of large language models for education and accessibility, with research internships at Microsoft Research and Amazon. I received my M.S. from UC Davis and B.S. from Zhejiang University.

News

  • Sep 2026 New paper: Fork Where the Model Changes Its Mind, on belief-shift branching for tree-structured reinforcement learning.
  • Aug 2026 Promoted to Senior Research Scientist at Salesforce AI Research.
  • Aug 2026 New paper: On the Fragility of Self-Improving Agents, examining variance, task order, and underspecification.
  • Jul 2026 New papers: Speculate with Memory on lossless acceleration for LLM agents, and Best Friends, Not Forever on persona collapse and behavioral drift in AI companions.
  • Oct 2025 Defended my Ph.D. in Computer Science at Columbia. Committee members: Prof. Julia Hirschberg, Prof. Kathleen McKeown, Dr. Michel Galley, Prof. Yang Liu, Prof. Zhou Yu.
  • Jul 2025 Started as a Research Scientist at Salesforce AI Research.

Publications

Teaching

  • 2023 COMS W4705: Natural Language Processing, Columbia University (Teaching Assistant).
  • 2022 COMS W4156: Advanced Software Engineering, Columbia University (Teaching Assistant).

Miscellaneous

Outside of work, I play basketball (peaked at twelve), chess (still learning to be good), and draw (no ChatGPT involved).