Ethan Elasky

About me

I am an AI researcher and co-founder of Palaestra Research, an independent nonprofit studying scalable oversight: the protocols, evaluations, and training methods that let weaker judges supervise stronger models. My work centers on AI debate, faithful reward labeling, legible model reasoning, and extending advances in problem solving beyond verifiable domains. I am currently supported by Coefficient Giving.

Previously, I was a Research Assistant at Academia Sinica (role in Chinese) and founding engineer at Acaceta. I studied data science, mathematics, and Chinese at UC Berkeley (magna cum laude, 3.98/4.00, Phi Beta Kappa).

I am a dual 🇺🇸 🇸🇰 citizen who speaks Mandarin fluently and Japanese at an intermediate level. In high school I was ranked top-50 worldwide in Lincoln-Douglas debate, and I authored Lincoln Douglas Debate from First Principles.

Outside of research, I enjoy reading, road biking, surfing, and playing pickleball, tennis, and golf.

Selected Projects

When does debate help a weak judge? Evidence from code and logic | 2026: Palaestra Research · Remote; With Frank Nakasako and Naman Goyal. When an independent critic improves a weak judge’s reward labels on code and logic tasks, and when it does not, across five model pairings.
Inference-time Generative Debates on Coding and Reasoning Tasks for Scalable Oversight | 2026: Palaestra Research · Remote; With Frank Nakasako. Generative debate versus consultancy on coding and reasoning tasks with verifiable ground truth; introduces the cleaned BigCodeBench+ benchmark. Supported by Coefficient Giving (formerly Open Philanthropy).
Quantifying Bias in Taiwanese Media | 2024: Advisor: Lucy Li · UC Berkeley; Honors thesis applying NLP techniques to measure bias in Taiwanese media coverage. Includes an interactive topic model visualization.
Lincoln Douglas Debate from First Principles | 2022: Independent · Berkeley, CA; Authored a textbook on Lincoln-Douglas debate that has helped hundreds of students improve at the activity.