What it is
A frozen LLM agent learns and revises explicit world models as natural-language rulebooks plus executable code, then plans with the verified model to improve itself through interaction. Instead of updating model weights, it stores revisable environment rules in external memory and compiles them to code. Prediction errors trigger rulebook and code revision, with updates accepted only if code matches the rules and exactly replays observed transitions.
Key results
- Clears every level of all 25 ARC-AGI-3 public games
- 100.0 mean RHAE on ARC-AGI-3, +59.3 vs Claude Opus 5 harness
- Atari Pong controller wins 21:0 in 3 episodes with no further LLM calls
Shows a frozen LLM can improve behavior by iteratively refining verified symbolic world models in external memory rather than updating parameters.