This repository is the project page for our experiments with coding agents designed to be game-general within ARC-AGI-3. We develop and evaluate agent architectures, reasoning strategies, and tools for solving unfamiliar interactive games.
The agents contain no game-specific code, prompts, hand-coded heuristics, hidden solutions, or other game-specific information. The same agent code and prompts are used across games.
Each game is evaluated as a single recorded playthrough from a fresh workspace. The agent has no access to files, logs, or conversation state from previous playthroughs. It cannot restart the game or return to completed levels.
With GPT-5.6-sol, ewma_sv_v1.6 fully solved all 25 public games at both tested reasoning efforts, achieved about 99% mean per-game Relative Human Action Efficiency (RHAE), and used fewer than half as many environment actions as the human baseline.
Because GPT-5.6-sol postdates the public games and held-out performance has not been tested, we interpret this as saturation of the public set—not evidence that ARC-AGI-3 has been solved generally.
For new experiments, consider:
ewma_sv_v1.6: maintains and simplifies a fixed-interface executable world model and verifies it through exact replay of recorded observations. It gives the strongest public-game results but is relatively expensive to run.twma_v1.6: maintains a textual world model without requiring a persistent executable simulator or replay-verification stack. It is much simpler and less expensive while still achieving strong results.
- Executable World Models for ARC-AGI-3 in the Era of Coding Agents — supporting materials
- Do Coding Agents Need Executable World Models, Simplification, and Verification to Solve ARC-AGI-3? — supporting materials
If you use this repository or the released results, please cite the relevant paper.
Released under the MIT License.