Huawei’s Noah’s Ark Lab, working with University College London and other teams, has introduced an AI agent system called Memento 3, according to Techub News. Built on Claude Opus 5, the system posted a perfect RHAE score of 100 across 25 games and 183 levels in the public ARC-AGI-3 set. It completed those tasks in 7,518 actions, about 56% fewer than humans needed for the same levels.
The team said Memento 3 uses a “code as model” approach. During gameplay, the agent writes and iterates on its own rulebook and an executable world model, allowing recursive self-improvement. The researchers also tested the system in Atari Pong. After discovering a control strategy, the AI no longer needed to call the large model and then won three straight games by 21:0.
According to the paper, the method does not require model weight updates or game-specific fine-tuning. Instead, it relies on an external world model so a general-purpose large model with frozen parameters can keep iterating in unfamiliar domains. The team added that experiments in some games showed a dual-world-model setup could cut the number of actions even more. The item cited QbitAI as the source of the paper details.
Huawei’s Noah’s Ark Lab, together with University College London and other teams, has launched an AI agent system called Memento 3, according to Techub News. The system is built on Claude Opus 5 and achieved a perfect RHAE score of 100 across 25 games and 183 levels in the public ARC-AGI-3 benchmark set.
It used 7,518 actions to complete those levels, about 56% fewer than humans needed to finish the same tasks.
Memento 3 uses a “code as model” design
The research team said the system adopts a “code as model” method, allowing the AI to write and revise its own rulebook and executable world model while playing. That setup enables recursive self-improvement during the task itself.
Three straight 21:0 wins in Atari Pong
The team also applied the system to Atari Pong. After the AI figured out a control strategy, it no longer needed to call the large model and went on to win three consecutive games by 21:0.
Paper says no weight updates or game-specific fine-tuning were needed
According to the paper, the method does not require updating model weights or fine-tuning for individual games. Instead, it builds an external world model so a general-purpose large model with frozen parameters can keep iterating on its own in unfamiliar domains.
The researchers also said tests in some games showed that using two world models could reduce the number of actions further. QbitAI was cited in the report.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
