Tsinghua Researchers Introduce Algebraic World Model for Exploration and Formal Proof

What happened

A Tsinghua research team led by Andrew Yao and Yang Yuan introduced Algebraic World Model (AWM), which turns observations into executable Python hypotheses, refines abstract rules, and formalizes claims in Lean 4. The article reports that AWM completed all 183 ARC-AGI-3 levels with a 99.9% score, using 43% of the human baseline’s average action count, and proved an altered level impossible within its step budget. Its proofs depend on whether exploration captures the relevant rules.

Read next

• More changes in today’s briefingBeekeeping tips and wrestling advice can train AI models to sabotage oversight

The story so far

A Tsinghua University team led by Andrew Yao and Yang Yuan introduced Algebraic World Model (AWM), a method for learning rules in interactive worlds that, according to QbitAI, completed all 183 ARC-AGI-3 levels with a 99.9% score while using 43% of the human baseline’s average action count. AWM turns observations into executable Python hypotheses, refines abstract rules through exploration, and formalizes claims in Lean 4. QbitAI also reported that it proved an altered level impossible to complete within that level’s step budget. The proofs depend on whether exploration captures the relevant rules, so their applicability to the interactive world rests on the rules the system learns.