Reproduce “Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jens Lehmann, Andrei Aioanei, Sahar Vahdati
A research group introduced Tycho, an agent employing active abstraction with programmatic world models to infer game mechanics and goals in interactive environments. The system achieved a 100.00 Relative Human Action Efficiency (RHAE) on all 25 ARC-AGI-3 games, completing all 183 levels with frontier language models and using 61% fewer actions compared to aggregate human baselines.
Key insights
- The 'Orchestrator' policy, which allows the main actor to delegate world model construction and repair, empirically outperformed both direct reasoning and automatic, trigger-based repair mechanisms.
- High transition accuracy in a world model does not directly translate to strong game-playing performance; effective goal inference, model allocation, and action selection are also critical.
- Programmatic world models, generated and debugged by large language models, provide an inspectable and effective representation for discrete, algorithmic interactive domains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.