Pairwise feedback
Chatbot Arena Ranking Algorithm
We study ranking and Pareto-set identification in Chatbot Arena through binary pairwise feedback, developing algorithms for both synthetic and realistic preference settings.
01 — Research
Questions I’m interested in.
R / 01
Reinforcement Learning
RL / Theory
RL Theory
R / 02
Large Language Model
LLM / Interpretability
LLM Interpretability
Mechanistic study
CoT for Counting Mechanism
We investigate how chain-of-thought supports counting by tracing the mechanism from controlled synthetic tasks to realistic needle-in-a-haystack settings.