Tsinghua's Andrew Yao team's new work: Algebraic World Model, enabling AI to think like Newton

What happened: A Tsinghua University team led by Andrew Yao and Yang Yuan introduced an Algebraic World Model (AWM) that turns observations into executable Python hypotheses, refines them into general rules, and formalizes them in Lean 4 for machine-checkable proofs. Key numbers: On ARC-AGI-3, AWM with Opus 5 cleared all 183 levels at 99.9% with average actions at 43% of the human Why it matters: The team says rule-recording models such as GPT Astra can assert wrong conclusions on impossible tasks, while AWM aims to state why a task is impossible and how close a feasible condition is.

A Tsinghua University team led by Andrew Yao and Yang Yuan introduced an Algebraic World Model (AWM) that turns observations into executable Python hypotheses, refines them into general rules, and formalizes them in Lean 4 for machine-checkable proofs.

The team says rule-recording models such as GPT Astra can assert wrong conclusions on impossible tasks, while AWM aims to state why a task is impossible and how close a feasible condition is.