An AI agent that writes its own rule book — in code — and updates it through trial and error just posted a strong result on a benchmark designed to resist exactly that kind of shortcut.
OPINE-World pairs two large language model agents in a loop: one acts inside the environment, the other translates observations into source code that models how the world behaves. When the code model is wrong, counterexamples from the acting agent trigger a rewrite — a technique borrowed from formal verification called counterexample-guided inductive synthesis. The system also tracks what it calls "ontology error," a Bayesian measure of whether it has correctly identified the types of objects in the scene, and steers exploration toward situations most likely to expose gaps in that understanding. The researchers tested it on ARC-AGI-3, a benchmark that deliberately withholds object definitions, goals, and action rules to force genuine skill acquisition.
The result — 20 of 25 games solved, with an action-efficiency score of 78.4 relative to the human baseline, and no per-game training — matters because it challenges the standard trade-off in world modeling. Deep-network world models are flexible but notoriously data-hungry and struggle outside their training distribution; program-synthesized models are efficient but have mostly been shown on tidy, structured environments where the object vocabulary is handed over in advance. OPINE-World attempts to get the efficiency benefits of code-based models without requiring that hand-holding.
ARC-AGI-3 is still a controlled benchmark, and five unsolved games out of 25 is not a rounding error — but for a system running without any per-game fine-tuning, it is a harder result to dismiss than the typical lab demo.