Mmastodon TechnologyTechnology first seen 4 d ago, last 4 d ago, peak #4

Chinese universities build agent that proves ARC puzzles unsolvable

Original: Tsinghua and Zhejiang University researchers have built an Algebraic World Model, an agent that plays ARC-AGI-3 puzzles

Researchers at Tsinghua and Zhejiang University have developed an Algebraic World Model, an AI agent that tackles ARC-AGI-3 puzzle levels. Formalised in the Lean 4 proof language, it not only plays the puzzles but can formally prove when a level cannot be won, reportedly showing all six sealed levels to be unwinnable where models like Claude Opus and GPT reportedly failed.

Why now: AI benchmark results showing formal mathematical proofs of unsolvability, outperforming major frontier models, are drawing attention in the AI research community

Tsinghua UniversityZhejiang UniversityARC-AGI-3Lean 4Claude Opus

Open on mastodon →

Evidence

• Tsinghua and Zhejiang University researchers have built an Algebraic World Model, an agent that plays ARC-AGI-3 puzzles and can formally prove a level cannot be won. Formalised in Lean 4, it proved all six sealed levels unwinnable, where 12 runs of Claude Opus 5.5 and GPT-6… · china@universeodon.com · 1

API: https://socialmediatrends-api.osmike.com/v1/trends/4707757