OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration
06:00 · July 3, 2026 · arXiv cs.AI RSS
Learning how an environment behaves from interaction is central to building agents that adapt to unfamiliar tasks. World models learned with deep networks are flexible but data-hungry and transfer poorly beyond their training distribution. Program-synthesized world models, written as source code by LLMs and refined through counterexample-guided inductive synthesis (CEGIS), are instead data-efficient and reusable, yet they have been demonstrated mainly on structured-state worlds with a given object vocabulary, and a single program search does not scale to pixel-rendered environments whose object structure must be hypothesized flexibly. We introduce OPINE-World, an LLM agent that learns an object-centric programmatic world model online from interaction. OPINE-World couples two cooperating agents in a loop of hypothesis and test, one acting in the environment and one synthesizing the model in code with replay verification and model-based planning, and it steers exploration with a Bayesian measure of object-type adequacy we call ontology error. We evaluate OPINE-World on ARC-AGI-3, a benchmark for skill-acquisition efficiency in which the object vocabulary, the goal, and the action semantics are withheld. OPINE-World solves 20 of 25 games without per-game training and reaches an action-efficiency score of 78.4 against the human baseline.
Summary
OPINE-World addresses the challenge of building agents that acquire reusable world models from interaction in environments where object structure, goals, and action semantics are unknown. Deep network-based world models offer flexibility yet require large amounts of data and generalize poorly outside their training distribution. In contrast, program-synthesized models expressed as source code can be data-efficient and inspectable, but prior systems have been limited to structured-state settings with a predefined object vocabulary and have struggled to scale to pixel-rendered environments.
The system couples two cooperating LLM agents in a continuous hypothesis-and-test loop over a shared replay buffer. One agent acts in the environment and maintains a natural-language description of observed dynamics, while the second synthesizes and refines an object-centric Python transition model through counterexample-guided inductive synthesis. Candidate programs are admitted only after they reproduce every recorded transition exactly. Exploration is guided by ontology error, a Bayesian measure of how well the current object-type partition explains observed behavior, allowing the agent to focus on objects whose dynamics remain poorly captured.
Once a verified model exists and at least one level has been cleared, a planner searches the model for goal-directed action sequences that are then validated step-by-step against the live environment. This architecture enables online discovery of both the object ontology and the transition rules without any per-game training or demonstrations.
Evaluated on the ARC-AGI-3 benchmark, which withholds object vocabularies, goals, and action meanings, OPINE-World solves 20 of 25 games and 160 of 183 levels. It achieves an action-efficiency score of 78.4 relative to the human baseline and outperforms both single-agent coding baselines and prior program-synthesis or neural latent-model approaches, which solve none of the games.
Why it matters
This research is highly relevant for Dutch AI researchers focusing on autonomous agents and explainable AI. The programmatic approach to world modeling aligns with the Netherlands' strategic emphasis on transparent, data-efficient, and interpretable AI systems.
