Memento 3 builds on prior Memento work by extending reflective learning from policy and procedural memory to explicit world models.
Frozen LLM agents learn world models via reflective rulebooks
Memento 3 maintains a natural-language rulebook, compiles it into executable code, and refines both through prediction errors, achieving perfect scores on ARC-AGI-3 and Atari Pong without updating the underlying model.
Chinese Tech
Haoyu Zhao · Zhengxu Yu · Zhiyuan He · Meng Fang · Rasul Tutunov · Haitham Bou-Ammar · +2 more
University College London · Huawei Noah's Ark Lab · University of Liverpool
Research Digest··3 min read
The authors introduce Memento 3, a system that enables a frozen large language model (LLM) to continually learn explicit world models by storing revisable hypotheses in an external natural-language rulebook.
Why this paper
From Huawei Noah's Ark Lab and 2 others
In one line
An agent learns explicit world models as rulebooks and code, clearing all ARC-AGI-3 levels with 44% of human actions.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§