Modeling Others' Minds as Code (arxiv.org)

Researchers released "Modeling Others' Minds as Code," a paper proposing ROTE, an algorithm that represents everyday social routines as executable behavioral programs rather than as policies tied to latent beliefs or rewards. ROTE uses large language models to synthesize a hypothesis space of small, human-readable programs (scripts like “wait for green, then go”) and applies probabilistic inference to maintain uncertainty and pick among program hypotheses from sparse observations. The authors evaluate ROTE in gridworld tasks and a large-scale embodied household simulator, showing it predicts human and agent behavior more accurately than behavior cloning and LLM-only approaches—improving in-sample accuracy and out-of-sample generalization by as much as ~50%. This program-synthesis framing matters because it yields compact, compositional, and interpretable models that are more sample efficient and adapt quicker than traditional data-hungry or computationally brittle belief-desire models. Key technical implications include leveraging LLMs as proposal/generation engines for structured behavioral hypotheses, then using probabilistic inference to reason under uncertainty and generalize to new contexts. That combination promises safer and more robust human-AI coordination in real-world settings by enabling faster, interpretable prediction and easier debugging of modeled behaviors.