Researchers from Huawei's Noah's Ark Lab, in collaboration with University College London and others, have developed Memento 3, an AI agent system capable of mastering 25 games from the ARC-AGI-3 benchmark. This system achieved a perfect score by using significantly fewer actions than human players, demonstrating a novel approach to AI self-improvement. Memento 3 utilizes a 'Code as Model' strategy, where the AI generates and refines its own rulebooks and executable world models, allowing it to learn and adapt without updating the underlying large language model, such as Claude Opus-5. AI
IMPACT This research demonstrates a novel method for AI self-improvement, potentially accelerating learning in complex environments without extensive model retraining.
RANK_REASON The item details a new AI agent system's performance on a specific benchmark, including its methodology and comparison to other systems. [lever_c_demoted from research: ic=1 ai=1.0]
- ARC AGI 3
- Atari
- baseline1
- Claude Opus-5
- Huawei
- Memento 3
- Nvidia
- Pong
- Princeton University
- University College London
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
