Huawei Noah's Memento 3 clears ARC-AGI-3 public set
华为AI靠RSI通关25款游戏！操作次数比人类少56%
C · Weak signal QbitAI news write-up relaying a Huawei Noah's Ark and UCL preprint (arXiv 2610.11794); it adds leaderboard context but no independent check, and the paper itself is the source to read.
Summary based on the full text·Original: about 6 minutes to read
In short
A QbitAI article (12 October 2026) on Memento 3, an agent system from Huawei Noah's Ark Lab with UCL and others, described in the paper "Model-Based Recursive Self-Improvement". Running on Claude Opus 5 with frozen weights and no game-specific fine-tuning, it cleared all 25 public ARC-AGI-3 games (183 levels) with a perfect relative human action efficiency score. The method, "Code as Model", has the agent keep a Markdown rulebook and a Python world model of each game, test them against real frames, and revise both under Git version control.
Key numbers
- 7518 vs 17135actions Memento 3 needed to clear all 25 ARC-AGI-3 public games versus the human total
- ~40.7average RHAE score of Claude Opus 5 in the official ARC Prize harness, against Memento 3's 100 on the same base model
- 899 to 597actions to clear game wa30 with one versus two parallel world models (about 34% fewer)
- ~9500game frames Memento 3 used to learn a Pong control policy that then won 21:0 three times without calling the model
As reported by the source.
Chinese·Web page·opens qbitai.com
What it reports
- Full score of 100 RHAE on the public set, with 7,518 actions versus 17,135 for humans (about 56% fewer) and 8,347 for the previous best entry, Sergey Rodionov's baseline1 at 99.
- Other entries cited: Princeton's Continual Harness 20.5, DreamTeam 38.1, OPINE-World 78.4, NVIDIA's NOOA 85.1.
- The world model is verified by replaying past actions and comparing the 64x64 frame cell by cell; in game tr87 the rulebook went through 11 versions.
- Ablation on three games: 617 actions with the rulebook versus 677 without (about 9% fewer actions, about 18% fewer agent turns).
- Two parallel world models with different rule hypotheses cut wa30 from 899 to 597 actions; tested on that game only.
- Pong: after about 9,500 frames the learned program played alone and won 21:0 in three starts; acting from screen frames without a world model scored -19, against -20.7 for random play.
How they got it
The article summarises the paper's results and figures. The underlying work runs Claude Opus 5 at extra-high reasoning effort inside the agent loop; the authors' own code generation and verification produce the world models.
Limits
- Secondary report; the journalist did not reproduce results and the paper was not read for this digest.
- Public-set results only; no private-set or cost figures beyond a note that token and compute use is high.
- Games have fairly clear rules; the article itself says transfer to messier real-world settings is unproven.
Why it matters for companies
The pattern is the interesting part for applied work: a frozen general model writes down rules, encodes them as executable code, checks the code against observed outcomes, and hands repeat execution to that code. That is a practical template for process automation where the expensive model learns the procedure once and cheaper deterministic code runs it. It also shows how much a harness changes a benchmark score for the same model, which matters when reading vendor leaderboards.
Key passages in translation
Chinese original
作为对比，Claude Opus 5在ARC Prize官方标准测试框架下，平均RHAE只有约40.7分。
By comparison, Claude Opus 5 in the ARC Prize official standard harness averages an RHAE of only about 40.7.
Chinese original
系统会重新模拟此前的操作，逐格对比64×64的游戏画面，确认代码能准确复现真实结果。
The system re-simulates earlier actions and compares the 64x64 game frame cell by cell to confirm the code reproduces the real result.
Chinese original
递归自我改进（RSI）并不一定要动模型底层权重。
Recursive self-improvement does not necessarily require touching the model's underlying weights.
Summary based on the full text. Read the original for the full argument and data.Read the original on qbitai.com