Hi, I’m Min Cai (蔡旻). I’m an M.S. student graduated from Shenzhen University, where I was supervised by Prof. Haodi Zhang. Before that, I obtained my B.A. in Translation from Beijing Language and Culture University. Currently, I’m interning at Zhipu AI, mentored by Dr. Dan Zhang. Meanwhile, I also work closely with Dr. Ziniu Hu, Dr. Shichang Zhang, and Dr. Difan Zou.

I have broad interests in ML and NLP, particularly in understanding the mechanisms behind neural language models (LMs), developing LLM agents capable of solving complex problems, and enhancing LLM reasoning abilities. Currently, my primary focus is on inference-time algorithms for alignment and reasoning in LLMs.

Specifically, my current research focuses on:

• Interpreting and controlling LLM behaviors for better alignment with human values (e.g., SelfControl)
• LLM Agents capable of solving complex tasks, such as multi-agent social deduction games(e.g., AvalonBench)
• Improving LLM reasoning abilities, particularly by introducing advanced inference-time algorithms like Monte Carlo tree search (e.g., Strategist), controlled text generation and representation engineering.

Selected Publications

AvalonBench: Evaluating LLMs Playing the Game of Avalon

Jonathan Light*, Min Cai*, Sheng Shen, Ziniu Hu (*equal contribution)

Code Paper Website

AvalonBench is a benchmark that explores the potential of Large Language Models (LLMs) Agents in playing the strategic social deduction game, Resistance Avalon.

Self-Control of LLM Behaviors by Compressing Suffix Gradient into Prefix Controller

Min Cai, Yuchen Zhang, Shichang Zhang, Fan Yin, Dan Zhang, Difan Zou, Yisong Yue, Ziniu Hu

Code Preprint Website

SelfControl is an inference time LLM control method that leverages LLM self-evaluation to control model behaviors through representation engineering.

Strategist: Learning Strategic Skills by LLMs via Bi-Level Tree Search (ICLR 2025)

Jonathan Light, Min Cai, Weiqin Chen, Guanzhi Wang, Xiusi Chen, Wei Cheng, Yisong Yue, Ziniu Hu

Code Preprint Website

Strategist is an advanced game agent that utilizes LLMs to acquire new skills for playing multi-agent games through a self-improvement process.