Research directions
Formal Reasoning
We build neuro-symbolic reasoning systems that combine learning and search, in domains including mathematical theorem proving (e.g., Minimo, ProofCompass), SAT (e.g., NeuroBranch), program generation (e.g., Synchromesh) and verification (e.g., dafny-annotator).
Abstraction & Representation Learning
We study how systems can create explicit, reusable symbolic abstractions from experience or data, such as tactics for theorem proving (Peano, TacMiner), skills in reinforcement learning (LEMMA and theoretical results), or programmatic representations for machine learning more broadly (LeaPR).
Open-Ended Discovery
We build open-ended systems that can pose their own problems and learn from solving them. These include mathematical conjectures (Minimo), new programs with formal specifications (Formal Disco, dafny-annotator), or puzzles for evaluating AI models themselves (The Token Games, MathCamps). We seek to understand what makes good problems and solutions, beyond correctness.