interactive-exploration

OPINE-World introduces an LLM agent that learns an object-centric programmatic world model online through interaction, using ontology-error-prioritized exploration and cooperating hypothesis-test agents, achieving strong results on ARC-AGI-3.