Algebraic World Model: AI That Discovers Laws Like Newton Through Formal Proof

Background: The Limits of Pattern-Matching AI

Current frontier models like GPT Astra excel at the ARC-AGI-3 benchmark (99.9% score) by extracting symbolic rules from game observations. However, these rules merely describe phenomena—like Tycho Brahe's precise planetary records—without capturing the underlying generative laws (Kepler's laws). When tested on a modified level (sc25) with an impossible step budget (18 steps vs. minimum 19), Astra hallucinated a false minimum of 22 steps. Across 12 such tests, top models (GPT-6 Astra, Claude Opus 5.5) correctly judged impossibility only 5 times, and 4 of those gave wrong step counts.

Algebraic World Model (AWM): From Exploration to Formal Proof

Developed by Tsinghua University's Yao Qizhi team, AWM addresses this by moving from pattern matching to mathematical modeling and verification. It operates in three stages:

Explore: Play the environment, write observed regularities as executable Python code (initial hypotheses).

Refine: Validate and generalize hypotheses into reusable algebraic objects and rules.

Prove: Formalize the refined rules in Lean 4, producing machine-checkable mathematical theorems.

This pipeline mirrors the scientific method: observation → hypothesis → experiment → abstraction → proof → new observation.

Category-Theoretic Foundations

AWM draws on category theory, specifically projective limits (defining new concepts by composing existing ones) and inductive limits (generalizing across instances to a common framework). These dual operations iteratively transform messy Python hypotheses into a concise, consistent, and provable world model.

Experimental Results

Impossibility proofs: On the unsolvable sc25 level, AWM (backed by Opus 5.5) required only 2 attempts to prove the level impossible under the given step limit and to identify that exactly one extra step would suffice.

Full benchmark: On the entire ARC-AGI-3 suite (183 levels), AWM with Opus 5 achieved 99.9% success, solving all levels with an average action count of just 43% of the human baseline—the highest efficiency of any known system.

Limitations and Outlook

AWM's proofs hold only for the "observed world"; incomplete exploration yields partial truths. The research emphasizes a continuous cycle of observation, hypothesis, experimentation, abstraction, and proof.

Team and Resources

Co-first authors: Pan Guancheng (Zhejiang University, visiting Tsinghua) and Tan Qixin (Tsinghua Yao Class). Advisors: Yuan Yang (Tsinghua IIAI, Shanghai Qizhi Institute) and Yao Qizhi (Turing Award 2000, Tsinghua IIAI/AI Institute Dean). Project page: https://ephemeral6.github.io/awm-site/; Paper: http://arxiv.org/abs/2610.40503.