AI System 'Twin' Masters ARC-AGI-3 Games via Test-Time World Models

A new AI system named Twin (Test-time World-model Inference) has demonstrated remarkable proficiency in solving ARC-AGI-3 grid games, a benchmark for continual learning. Developed by researchers, Twin uses a frontier coding agent to write an executable world model for each game, inferring rules and goals purely from simulation and interaction, without hand-engineering. The system's inductive prior over grid games enables it to recover true transitions and goals on nearly all levels. Twin employs a twin world model for replay validation, ensuring actions are only taken after the program reproduces all previous observed transitions. Mismatches between predictions and actual results serve as counterexamples to repair the model. Twin successfully cleared 179 out of 183 levels (97.8%), outperforming humans in efficiency on 158 of those levels (88.3%). This approach contrasts with traditional methods that require custom designs per task, highlighting a shift toward test-time world-model inference for continual learning.