My research addresses sequential decision making under partial observability, delayed feedback, and confounded observational data, with an emphasis on settings where actions are not reversible. Application areas include clinical decision making, automated scientific discovery, and foundation model reasoning.
The four areas below list the active themes within each, along with the prior work they build on.
Methodological work on offline RL and sequential decision making: irreversibility, credit assignment under delay, uncertainty under partial observability, and structured action spaces.
Offline sequential decision making on observational clinical data: dead-end identification, treatment disparities, transfer across populations and sites, and uncertainty over irregular observations.
Reinforcement learning for foundation model reasoning: cross-domain generalization of RL recipes, compute allocation, post-training dynamics, distillation, and structured actions in agentic settings.
Sequential decision making for automated experimentation: embodied scientific reasoning, acting while prior experiments remain unresolved, and synthetic data for underrepresented situations.
A complete list is available on the Publications page. For inquiries about collaboration or student positions, get in touch.