Tsinghua's Algebraic World Model solves all 183 ARC-AGI-3 levels at 99.9%, proving unsolvable ones
量子位 · wechat · 2026-11-01
A team led by Turing Award winner Andrew Yao (Yao Chi-Chih) and Yuan Yang at Tsinghua proposed the Algebraic World Model (AWM), teaching AI to reason about world rules rather than just record phenomena.
Motivation: On rule-free ARC-AGI-3 games, top models like GPT Astra hit 99.9% but only produce "Tycho-style" observational notes. In a modified unsolvable-level test across 12 trials, GPT-6 Astra and Claude Opus 5.5 combined detected impossibility only 5 times, usually with wrong step estimates.
Method: AWM works in three steps — write observations as executable Python hypotheses; refine them into general objects and rules with new observations; formalize into machine-checkable Lean 4 proofs. It draws on category theory's inverse and direct limits.
Results: On the unsolvable level, AWM with Opus 5.5 proved impossibility in just 2 attempts and noted one extra step would suffice. It cleared all 183 ARC-AGI-3 levels at 99.9% with only 43% of the human baseline's action count.
Co-first authors are undergraduates Pan Guancheng (Zhejiang) and Tan Qixin (Yao Class, Tsinghua).
More from Models
- How LLM inference actually works: from 140GB weights to GPUs and model serving, explained as a story — _jaydeepkarale · 2026-11-04
- $100 of GPU time edits model weights to cut thinking tokens by 40% — jarrodwatts · 2026-11-04
- Sarvam open-sources Prism, a 44ms decision model that skips text entirely — heyshrutimishra · 2026-11-04
- Frontier models' teaching ability is 'off the charts', says dev: your 'I know kung fu' moment — shauseth · 2026-11-04
- kalomaze: crude diffusion autoencoders may lose low-frequency info at the latent stage — kalomaze · 2026-11-04
- Dev complains frontier models can't do light-touch English editing — A_K_Nain · 2026-11-04