Elizabeth Donoway
Bio
Elizabeth Donoway is an Anthropic Fellow and physics PhD candidate at UC Berkeley, supported by NSF Graduate Research Fellowship Program and Ford Foundation Predoctoral Fellowship awards. She works with Jan Leike at Anthropic on quantifying elicitation and learning in large language models, using information-theoretic approaches combined with interventions on model internals to understand the distinction between eliciting latent capabilities and teaching entirely new skills. She also participated in the MATS (ML Alignment & Theory Scholars) summer 2024 cohort under Marius Hobbhahn, where she developed evaluations of long-horizon goal-directedness in LLMs and LLM agents, and contributed to BIG Bench evaluations of physical reasoning in language models. Her PhD research in Mike DeWeese's group applies techniques from statistical mechanics and information geometry to understand how machines learn. Prior to her AI-focused work, she conducted condensed matter physics research using novel optical techniques to study exotic phases of matter and emergent phenomena in quantum materials.
Links
- Personal Website
- https://elizabethdonoway.com/
- Twitter / X
- LessWrong
- -
Grants
No grants recorded.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.
Details
- Last Updated
- Mar 22, 2026, 3:55 PM UTC
- Created
- Mar 20, 2026, 3:01 AM UTC