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

arXiv 2026 63.3 method, system

TLDR

A coding agent uses an executable Python world model to solve ARC-AGI-3 games, achieving 15/25 full solves with GPT-5.5.

Reasoning

The paper presents a novel approach combining executable world models with coding agents, showing strong results on public ARC-AGI-3 games. However, it lacks private validation and relies on specific LLM versions, limiting generalizability.

Read-first score

Read-first score 63.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark,result,validation

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=artifact,code,github

Citation impact 18%

74.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.74303964

Topical relevance 29%

45.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierBridgeMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 520.

Keyword Scores

world model

world simulator

world dynamics prediction

interactive world model

generative world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Executable Python world model that is verified against previous observations
Refactoring toward simpler abstractions as a practical proxy for an MDL-like simplicity bias
Planning through the model before acting
Scripted controller with predefined world-model interfaces, verifier programs, and plan executor, containing no game-specific logic
Same agent-facing prompts, workspace, and controller used across all games, with no hand-coded game-specific information
Auditing and closing unintended information channels in the harness to reduce benchmark-specific information exposure

Methodology

The system uses a coding agent that maintains an executable Python world model, verifies it against observations, refactors it toward simpler abstractions (as a proxy for MDL simplicity bias), and plans through the model before acting. It employs a scripted controller, predefined world-model interfaces, verifier programs, and a plan executor, with no hand-coded game-specific logic. The same agent and prompts are used across all games, and each playthrough starts from a fresh agent instance and clean workspace with no access to prior state.

Key Results

With GPT-5.5 high reasoning effort, the agent fully solved 15 out of 25 public ARC-AGI-3 games and achieved a mean per-game RHAE of 58.12%. With GPT-5.4 high reasoning effort, it fully solved 8 games and achieved a mean per-game RHAE of 41.29%.

Limitations

Performance on the private validation set, which is not yet available, remains to be tested
Results are only preliminary evidence and based on a limited set of 25 public games
The system relies on specific LLM versions (GPT-5.5 and GPT-5.4) with high reasoning effort, which may not be widely accessible or generalizable
The approach is intentionally direct and may have limitations in handling more complex or diverse ARC-AGI-3 tasks
Earlier harnesses had unintended information channels that required auditing and closure, indicating potential fragility in system design

Resources

Paper Code