Executable World Models for ARC-AGI-3 in the Era of Coding Agents

Researchers demonstrate a coding-agent system for ARC-AGI-3 that uses executable Python world models to solve abstract reasoning challenges without game-specific code. The agent achieved full solutions on 7 of 25 public games, establishing a generalizable baseline approach that relies on model verification and simplicity-driven refactoring rather than hand-coded logic.

This research introduces a practical framework for autonomous problem-solving that diverges from traditional hardcoded game engines. Rather than programming specific solutions for each challenge, the agent constructs executable world models dynamically, validates them against observations, and iteratively simplifies them toward abstract principles. This approach mirrors human reasoning patterns—building mental models, testing them against reality, and distilling core patterns.

The methodology reflects broader progress in agentic AI systems that prioritize generalization over task-specific optimization. By eliminating hand-coded logic, the researchers create a reproducible baseline suitable for evaluating future improvements. The 32.58% mean per-game Relative Human Action Efficiency demonstrates meaningful but incomplete performance, indicating clear room for advancement without suggesting premature capability claims.

For the AI development community, this work provides both a benchmark and architectural template for building reasoning systems that transfer across domains. The emphasis on model verification and simplicity bias connects to theoretical work on minimum description length (MDL) principles, grounding the practical engineering in formal foundations. This matters because production AI systems increasingly require explainability and robustness—verifiable world models address both concerns simultaneously.

The framework's success on solving 7 complete games without domain knowledge suggests scaling this approach could yield significant improvements. Performance testing on private validation sets will clarify whether public-set results generalize or reflect overfitting to available examples. Key variables to monitor include computational efficiency, scalability to more complex environments, and whether model interpretability translates to usable explanations for human operators.

• →Agent solved 7 of 25 ARC-AGI-3 games using generalizable executable world models without task-specific code.
• →Verifier-driven architecture validates and refactors models toward simpler abstractions, mimicking principled reasoning approaches.
• →32.58% mean per-game efficiency establishes a game-agnostic baseline for evaluating future coding-agent improvements.
• →Framework prioritizes interpretability and robustness through explicit model verification rather than black-box optimization.
• →Private validation results remain pending; public-set performance generalization will determine practical applicability.