How to Have a Sensitive Debate: An Instance-Optimal Protocol for AI Debate
A simpler and better debate protocol for scalable oversight
Jiawei Li*, Zhiyang Xun*, Lijie Chen, Jonah Brown-Cohen
I'm a final-year Ph.D. student in Computer Science at UT Austin, advised by Eric Price and David Zuckerman.
I received my B.S. from the ACM Honors Class at Shanghai Jiao Tong University.
Research interests
As language models take on more complex tasks and get better at reward hacking, they become increasingly difficult for humans to directly monitor or for programs to reliably judge. I develop methods to scale our supervision, in both training and evaluation:
- Scalable oversight via debate;
- Preventing reward hacking in RL;
- Scalable measurement of model behavior and misalignment.
My background is in theoretical computer science, and I still enjoy algorithmic questions beyond LLMs, such as those in diffusion models and complexity theory.