Scholar
Julian Michael
Google Scholar ID: 9DDOHR8AAAAJ
Scale AI
Citations & Impact
Citations
15,462
H-index
21
i10-index
27
Publications
20
Co-authors
83
list available
Contact
Publications
Resume (English only)
Academic Achievements
- Published 'GPQA: A Graduate-Level Google-Proof Q&A Benchmark' (COLM 2024, Spotlight), introducing a challenging QA benchmark.
- Co-authored 'Debate Helps Supervise Unreliable Experts' (arXiv preprint), exploring debate as a supervision method for unreliable experts.
- Received Best Paper Award at The Big Picture Workshop (2023) for 'The Case for Scalable, Data-Driven Theory'.
- Published 'Inducing Semantic Roles Without Syntax' (Findings of ACL 2021) with Luke Zettlemoyer.
- Released the GPQA dataset to push the boundaries of difficult evaluations for scalable oversight.
- Delivered invited talks at institutions including NYU and UW on scientific paradigms in NLP and AI ethics.
Background
- Currently works on AI safety, evaluation, and alignment at Meta.
- Research interests include AI alignment, formal semantics of natural language, scalable oversight, agent alignment, and using debate as a training and evaluation paradigm.
- Aims to advance the science of language through data and machine learning, especially in syntax and semantics.
- Advocates for empirical methods in the Science of AI and NLP to better understand intelligent behavior and language use.
- Formerly led the Safety, Evaluations, and Alignment Lab (SEAL) at Scale AI.
