Published in Preprint. In-review, 2026
LLMs can rewrite their own skills. But how good are they at learning from each other? We study recursive social improvement and find that copying useful discoveries can save learning effort, yet comes at the expense of independent exploration.
Download here
Published in Accepted Conference Paper @ NeurIPS 2026, 2026
Can self-interested, self-improving, and self-replicating agents learn to cooperate? We show that when social behavior, computation, and reproduction share one energy budget, cooperation can evolve from scratch.
Download here
Published in Accepted Conference Paper @ ICLR 2026, Oral Paper @ CogSci 2026, Best Paper award @ NeurIPS 2025 LAW Workshop, 2025
How can AI quickly and accurately predict the behaviors of others? We show an AI which uses Large Language Models to synthesize agent behavior into Python programs, then Bayesian Inference to reason about its uncertainty, can effectively and efficiently predict human actions.
Download here
Published in Oral Paper @ ICML, Presenting abstract @ CogSci, 2025
How can AI develop the ability to cooperate with novel people on novel problems? We show AI learning to cooperate in “self-play” with one partner on many environments helps agents meta-learn to cooperate with humans and other AI on problems they have never seen before.
Download here
Published in IROS, 2024
How can we build robots that intelligently and proactively communicate with humans? When and what should the robot communicate with the human user? In GOMA, we formulate communication as the minimization of the divergence of agents’ beliefs and plans.
Download here
Published in AAAI, 2024
How can agents efficiently perform recursive Theory of Mind reasoning? We propose a neurosymbolic method which approximates the posterior belief distribution through bottom up neural network proposals, then performs top down Bayesian inference over the most likely beliefs.
Download here