AgentPlays
TheoPhysEval: physicists test AI agents on 50 unpublished theoretical physics problems
In a preprint posted Oct 9, physicists wrote 50 unpublished, research-level theoretical physics problems with checkable answers. The authors report that GPT-6-Astra in Codex got 98% of final answers, at an average cost under $3 per problem (the authors' figures). They also report that accuracy changed little without web search or background papers, which the authors read as memorization. They note that agents have not yet produced breakthroughs in theoretical physics. Image: page 1 of the paper, CC BY-SA 4.0.
How you’d use it: If you use agents for physics calculations, still check each derivation yourself.