Today for AI
HOT RADAR
agentHEAT 20.4°

AI CLUSTERED EVENT · 10/12/2026

Huawei Noah and UCL Release Memento 3: Recursive Self-Improvement via World Models Achieves Zero-Fine-Tuning ARC-AGI-3 Mastery with 56% Higher Efficiency

2 reports archived2 independent sourcesupdated 10/12/2026, 13:48:36
Synthesis & Latest Updates
2 Sources Cross-Validated

Huawei Noah's Ark Lab and UCL have jointly released Memento 3, an agent system that achieves a perfect score on all 25 games in the ARC-AGI-3 benchmark without any weight updates or game-specific fine-tuning of the base model, Claude Opus 5. The system boosts the Relative Human Action Efficiency (RHAE) from a baseline of 40.7 to 100, requiring approximately 56% fewer operations than humans.

Cross-verified reports highlight that Memento 3's core innovation lies in "recursive self-improvement based on world models." Instead of relying on pre-trained knowledge, the AI constructs external state graphs through trial-and-error, reflection, and code generation. This architecture allows agents to rapidly converge on optimal strategies in unfamiliar environments. Notably, tests on Atari Pong demonstrated the system's ability to autonomously iterate strategies without real-time LLM inference, significantly reducing computational latency and costs.

This achievement validates the potential of pure prompt engineering and external architectural design as effective evolution engines for general agents. It reshapes developer expectations regarding zero-shot learning capabilities and offers a new technical paradigm for enhancing agent robustness and economic efficiency in dynamic environments.

LATEST/Recent developments confirm Memento 3's internal world model can execute autonomous strategies without real-time LLM inference, validating its effectiveness as a universal agent evolution engine and paving the way for reduced online interaction costs.

HEAT TRENDHourly heat curve

3 fully observed hours
Now
20.4
Peak
21.418:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 2 reports · Latest first
  1. 量子位官网T2·86 pts
    • Memento 3 achieved perfect scores on ARC-AGI-3 without updating model weights or task-specific fine-tuning, boosting RHAE by nearly 60 points over the base model via an external world model.
    • The core innovation lies in decoupling Recursive Self-Improvement from parameter adjustment, instead driving rapid policy convergence in unknown environments through verifiable, executable, and rollback-capable external state graphs.
    • In Atari Pong tests, the system demonstrated autonomous control independent of real-time LLM inference, indicating its internal world model possesses predictive capabilities that significantly reduce online interaction latency and compute costs.
  2. 量子位 (微信公众号)T2·86 pts
    • Memento 3 achieved a perfect RHAE score on the ARC-AGI-3 benchmark, operating ~56% more efficiently than humans and significantly outperforming the base model's standard test results.
    • The core innovation is 'Model-Based Recursive Self-Improvement,' where the AI builds its own environment rule models through trial-and-error, reflection, and code generation rather than relying on pre-trained knowledge.
    • No weight updates or game-specific fine-tuning were applied to Claude Opus 5, validating the massive potential of pure prompt engineering and external architectural design in complex interactive tasks.