TWIN Lets Coding Agents Reconstruct Unknown Games in Real Time, Raising ARC-AGI-3 Score from 7.8 to 93.3

TWIN reframes ARC-AGI-3’s interactive tasks as “reconstruct the environment first, then plan actions.” The agent starts with blank `step(grid, action)` and `goal_reached(grid)` functions, modifying the Python program based on each observed transition of the 64×64 grid. An execution gate validates the model cell by cell against the entire interaction history; if even one cell differs, it refuses to submit the next scored action. The planner then uses BFS to search for paths inside this executable replica. When the prediction first diverges from the real environment, the system stops immediately and uses the result as a counterexample to patch the model again.

The [paper](https://arxiv.org/abs/2608.14490) reports that the system completed 23 of 25 public games and 179 of 183 levels, achieving a score of 93.3. The same base model scored only 7.8 when interacting directly and 61.1 when used with a general-purpose agent harness. In 156 of the completed levels, the system inferred the correct goal before receiving its first reward. The researchers also retrospectively tested actions that had not actually been taken. The final world models achieved a 68.6% exact-frame accuracy, indicating that they did more than merely replay past interactions, although they remained far from accurate simulators.

The technical value lies in converting expensive, irreversible environment interactions into locally executable program validation that can be repeated indefinitely. This pattern could extend to browser agents, robotics, or infrastructure operations. However, the results rely heavily on ARC-AGI-3’s rule that internal computation is free while real actions count toward the score, and the system consumed approximately 2.6 billion inference tokens. The publicly available [code and execution artifacts](https://github.com/Alexyskoutnev/TWIN-ARC-AGI-3) allow others to inspect the records, but each game currently has only one official run—too little evidence to estimate stability across models. The next question is whether a strict replay gate can still work in environments that are partially observable, stochastic, or constrained by simulation costs.