Huawei AI passes 25 games using RSI! 56% fewer operations than humans

What happened: Huawei Noah's Ark Lab and UCL teams built Memento 3, an AI agent that cleared all 25 games and 183 levels in the ARC-AGI-3 public set using a method called Model-Based Recursive Self-Improvement. Key numbers: 7518 operations versus 17135 for humans, about 56% fewer. RHAE score 100. Claude Opus 5 baseline scored about 40.7. Why it matters: The system improved results without updating model weights or fine-tuning, and later ran Pong without calling the large model. The team notes costs from high reasoning effort and that validation is limited to rule-clear game environments.

Huawei Noah's Ark Lab and UCL teams built Memento 3, an AI agent that cleared all 25 games and 183 levels in the ARC-AGI-3 public set using a method called Model-Based Recursive Self-Improvement.

The system improved results without updating model weights or fine-tuning, and later ran Pong without calling the large model. The team notes costs from high reasoning effort and that validation is limited to rule-clear game environments.