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

2026-08-14

AI Summary

This study addresses the challenges of continual learning and goal inference in games with unknown rules by proposing a test-time world model reasoning framework. The approach leverages code agents to construct executable digital twins that autonomously infer game rules and objectives without reward signals. Model bias is mitigated through replay verification and counterexample repair mechanisms, enabling efficient continual adaptation. Experimental results on benchmark tasks demonstrate that the system achieves a 97.8% completion rate, surpasses human efficiency by 88.3%, and attains a score improvement of 93.3%. These outcomes significantly outperform existing methods, validating the framework’s effectiveness in open-ended environments where prior knowledge of game dynamics is unavailable.

Abstract

We present a Test-time World-model Inference (Twin) system, in which a frontier coding agent writes an executable world model for completing continual learning tasks, such as ARC-AGI-3 games. Traditional approaches hand-engineer such models, one custom design per task. Each game hides its rules and goal, and our system constructs them from simulation and interaction alone. Its inductive prior over grid games is strong enough to recover the true transitions of the game and the goal on nearly all levels. Replay validation happens in a twin world model. The harness enforces that an action is not made until the program reproduces every previous observed game transition. Each mismatch between a world model prediction and the actual action result becomes a counterexample that is used to repair the world model. Twin clears 179 out of 183 levels (97.8%), and does so more efficiently than humans in 158 out of 179 levels (88.3%). The system infers the goal before any reward on 156 of the levels it clears (87.2%), and in the remaining levels automatically discovers the goal by search. The benchmark scores completion and action efficiency, between 0 and 100, against humans playing each game for the first time. Played directly, the base model scores only 7.8%; an off-the-shelf harness increases it to 61.1%, whereas our twin world model increases the same base model to 93.3%, clearing 23 out of 25 games. Building a usable world model is simpler than anticipated, whereas the harder problem is inferring the right goal.

Problem

• Unknown Game
• World Model Inference
• Continual Learning
• Goal Discovery
• Test-Time Adaptation

Innovation

• Test-time World-model Inference
• Executable World Model
• Replay Validation
• Goal Inference
• Continual Learning