OPINE-World Learns Programmatic World Models from Interaction

Building adaptive agents that can learn environmental behaviors from interaction is a core challenge in AI. While deep network world models are flexible, they are data-intensive and struggle to transfer beyond their training distribution. Program-synthesized world models, generated by LLMs and refined through counterexample-guided inductive synthesis (CEGIS), offer data efficiency and reusability but have been limited to structured-state worlds with predefined object vocabularies. OPINE-World introduces an LLM agent that learns object-centric programmatic world models online through interaction. It features two cooperating agents: one acts in the environment, and the other synthesizes the model in code, using replay verification and model-based planning. The system intelligently guides its exploration using a Bayesian measure called "ontology error," which assesses the adequacy of object types. Evaluated on ARC-AGI-3, a benchmark for skill acquisition efficiency, OPINE-World solved 20 out of 25 games without per-game training, achieving an action-efficiency score of 78.4 against the human baseline, demonstrating its ability to adapt to environments where object vocabulary, goals, and action semantics are unknown.