Twin Clears 97.8% of ARC-AGI-3 Levels by Making a Coding Agent Write an Executable World Model — Base Model Alone Scores 7.8%

arXiv 2608.14490·high signalXBlueskyLinkedInCopy link

Twin (Test-time World-model Inference) has a frontier coding agent write an executable simulator of an unknown grid game, and a harness refuses to let the agent act until that program reproduces every previously observed transition; each mismatch becomes a counterexample used to repair the model. It clears 179 of 183 levels (97.8%), beats first-time human action efficiency on 158 of 179, and lifts the same base model from 7.8% played directly and 61.1% under an off-the-shelf harness to 93.3%. The authors' conclusion is the builder-relevant part: constructing a usable world model turned out easy, while inferring the goal was the hard problem — it was inferred pre-reward on only 87.2% of cleared levels.