Twin: Playing an Unknown Game with a Test-Time Digital Twin
Twin has a coding agent build and continuously repair an executable simulator before acting. The harness lifts the same base model from 7.8 to 93.3 on ARC-AGI-3 by validating every observed transition.
Twin makes a coding agent infer an executable world model from interaction. Before any action, the harness requires that model to reproduce every observed transition; mismatches become counterexamples for repair. It clears **179 of 183 levels**.
For agents operating in unfamiliar stateful systems, add an executable hypothesis layer and replay gate instead of relying only on verbal reasoning. Treat each real action result as a test that must pass before the agent proceeds.
Twin makes a coding agent infer an executable world model from interaction. Before any action, the harness requires that model to reproduce every observed transition; mismatches become counterexamples for repair. It clears **179 of 183 levels**. For agents operating in unfamiliar stateful systems, add an executable hypothesis layer and replay gate instead of relying only on verbal reasoning. Treat each real action result as a test that must pass before the agent proceeds. On the 25-game benchmark, the same base model scores **7.8** directly, **61.1** with an off-the-shelf harness, and **93.3** with Twin. These are grid-game results with a strong domain prior; the authors say goal inference remains harder than modeling transitions.
Twin turns runtime validation into an executable hypothesis test: the agent must model and replay every observed transition before taking another real action. This supplies a stronger, domain-specific form of harness gating than response checks alone and demonstrates large gains in replayable grid games. It also narrows applicability: unfamiliar production systems may lack cheap replay, and correct transition modeling does not solve goal inference.