Tsinghua team's Algebraic World Model turns game exploration into Python rules and Lean 4 proofs
Overview
A Tsinghua University team led by Yao Qizhi and Yuan Yang proposes the Algebraic World Model (AWM), which turns an AI's game exploration into executable Python hypotheses, refines them, and formalizes them in Lean 4.
QbitAI reports that on the ARC-AGI-3 test set, AWM with Opus 5 cleared all 183 levels at a 99.9% score, using 43% of the human baseline's average actions. The article also says AWM proved that one altered level was unsolvable at its step limit, a case where GPT-6 Astra and Claude Opus 5.5 gave wrong answers. These results are the article's account and have not been independently verified.
Written by AI from the articles below · updated Nov 1, 6:35 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- QbitAINewsTsinghua team's Algebraic World Model lets AI reason about rules from experiments
AITsinghua University researchers led by Yao Qizhi and Yuan Yang present the Algebraic World Model (AWM), which turns AI game exploration into executable Python hypotheses, refines them, and formalizes them in Lean 4. On the ARC-AGI-3 test set, AWM reports clearing all 183 levels with Opus 5 at a 99.9% score, using 43% of the human baseline's average actions. The article says the model also proved that one altered level was unsolvable at its step limit, where GPT-6 Astra and Claude Opus 5.5 gave wrong answers.
Heat Trend
- Current comparable heat
- 4
- Peak comparable heat
- 9Nov 1, 7:00 AM ET
- Change in past 24 hours
- -49%
The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.