Huawei and UCL’s Memento 3 Clears 25 ARC-AGI-3 Games Without Updating Model Weights

What happened

Huawei’s Noah’s Ark Lab, UCL, and collaborators report that Memento 3 cleared all 25 public ARC-AGI-3 games and 183 levels using 7518 actions, about 56% fewer than humans, achieving a 100 RHAE score. Built on Claude Opus 5 without weight updates or game-specific fine-tuning, the agent iteratively revises a Markdown rulebook and executable Python world model. In Atari Pong, its learned code-based strategy won three games 21:0 without further LLM calls; the experiments primarily cover game environments with relatively clear rules.

The story so far

Huawei’s Noah’s Ark Lab, UCL, and collaborators introduced Memento 3, an AI agent they report cleared all 25 public ARC-AGI-3 games and their 183 levels without updating model weights or using game-specific fine-tuning. The team reported that it used 7518 actions, about 56% fewer than humans, and achieved a 100 RHAE score. Built on Claude Opus 5, Memento 3 iteratively revises a Markdown rulebook and an executable Python world model. Its improvements come through those revisions rather than changes to the underlying model’s weights. In a separate Atari Pong experiment, the researchers reported that its learned code-based strategy won three games 21:0 without further calls to the language model. The experiments primarily cover games with relatively clear rules. Claude Opus 5 ran at Extra-High reasoning intensity, and exploration, code generation, and verification still required substantial tokens and computing resources. Whether the approach works in more complex, changing real-world environments remains untested.

Keep reading

Trending now

1. 1Meta Releases Muse Spark Open With Mac and Cloud Support, Topping US Open-Weight Index
2. 2Google Launches Gemini 4 Argon Globally, Expands Paid Access as Benchmark Results Vary
3. 3OpenAI Updates GPT-6 Astra Pro in ChatGPT With Parallel Agents and Faster Responses
4. 4Meta’s New Business Terms Raise Dispute Over Website Crawling and AI Training Rights