Twin: Playing an Unknown Game with a Test-Time Digital Twin

TL;DR - Twin uses a coding agent to construct and continually repair executable world models for unknown ARC-AGI-3 games at test time. This raises the base model’s benchmark score from 7.8% to 93.3%, showing that validated environment modeling can substantially improve agent performance.

• Twin requires its model to reproduce every observed transition before taking another action.
• Prediction mismatches become counterexamples used to repair the executable world model.
• It clears 179 of 183 levels and beats human action efficiency on 158 of the 179 cleared levels.
• It infers the goal before receiving any reward on 156 cleared levels; goal inference remains harder than modeling transitions.