Twin agent clears 179 of 183 ARC-AGI-3 levels

Twin cleared 179 of 183 ARC-AGI-3 levels by writing a working simulator before it acted, per a new arXiv paper. For anyone building agents in unfamiliar environments, that's the failure point worth watching: the world model, not the acting. On the same base model, Twin's loop lifted scores from 7.8% played directly to 93.3%, up from 61.1% under a standard harness.

The system won't let the agent touch the game until its program matches every transition already observed. Each mismatch becomes a counterexample, which the agent uses to rewrite the model. That loop repeats until the simulation matches the real game, and only then can the agent act inside it.

That approach cleared 179 of 183 levels, a 97.8% rate, and beat first-time human action efficiency on 158 of them, according to the paper.

The paper's own conclusion cuts against the headline number. Building a working simulator turned out to be the easy part. Inferring the actual goal of each game was hard. The system worked out the goal before earning any reward on only 87.2% of the levels it cleared.