Tsinghua and Zhejiang's 'Algebraic World Model' formalizes ARC-AGI-3 game rules in Lean 4, proving six sealed levels are 'unsolvable' (report)

Verified (fact)

According to Pandaily, the Algebraic World Model (AWM) from researchers at Tsinghua University and Zhejiang University learns the rules of ARC-AGI-3 games on its own and then formalizes them in Lean 4. Using this approach, it proved that all six sealed levels are structured so that they cannot be cleared. According to the report, frontier models failed on these levels.

Needs fact-check (claim)

Proved that all six sealed levels on which frontier models failed are impossible to clear

Why it matters (interpretation)

This is a case of combining a world model with formal proof to verify the validity of the benchmark tasks themselves. It could feed into discussion of the reliability of ARC-AGI-3 evaluations.

Sources

• Press Tsinghua's Algebraic World Model Proves ARC-AGI-3 Levels Unwinnable With Lean 4 Pandaily

More about Tsinghua University

• 2026-10-31 P2 Research Paper AgentsEfficiency & infra Zhipu AITsinghua University
  Tsinghua and Zhipu's 'HarnessPro' lets agents rewrite their own harness from execution logs, claims 26.1% lower cost per task (report)

  Researchers from Tsinghua University, Zhipu AI and others introduced 'HarnessPro.' In this approach, an agent analyzes its own execution...

• 2026-10-30 P2 Research Unconfirmed Robotics & physical AIReinforcement learning Tsinghua UniversityInfinigence AIAgiBot
  Tsinghua, Infinigence AI and AgiBot present method for 96 real robots to learn from each other's experience; newly added robots reach 97.5% success (report)

  According to Pandaily and QbitAI (量子位), a joint team from Tsinghua University, Infinigence AI and AgiBot introduced a method (Collective...

• 2026-10-28 P3 Research Paper Robotics & physical AIComputer vision Tsinghua University
  Tsinghua develops 588g autonomous liver ultrasound robot that learned doctors' probe handling, evaluated on 119 people (report)

  According to AI Times Korea (인공지능신문), researchers at Tsinghua University's School of Biomedical Engineering trained AI on skilled...

• 2026-10-18 P2 Open source Generally available AgentsAI platforms & solutions Tsinghua UniversityModelBestOpenBMB
  Tsinghua's THUNLP, ModelBest and OpenBMB release open-source agent OS 'PilotDeck 2.0', with directly viewable and editable memory and lightweight models for easy subtasks (report)

  According to Pandaily, Tsinghua University's THUNLP, ModelBest and OpenBMB have released 'PilotDeck 2.0', an open-source agent operating...