AI Agent Solves 49/50 Unpublished Research-Level Theoretical Physics Problems
The paper introduces TheoPhysEval, a benchmark of 50 unpublished, research-level theoretical physics problems with verifiable answers, each solvable by a world expert in about a day or less. Testing off-the-shelf AI agents without human hints, the best performer—GPT-6-Astra run in Codex with tool access (literature search, scripts for symbolic/numerical work)—achieved 98% accuracy at under $3 per problem, and accuracy barely dropped when web search was disabled, suggesting heavy reliance on memorized literature. The authors note this raises the question of why such strong problem-solving hasn't yet translated into AI-led research breakthroughs in theoretical physics. Twitter discussion centered on astonishment at the near-perfect score and reports that some answers initially marked wrong were actually correct—meaning the human-generated answer key itself contained errors. Commenters (including a theoretical physicist reportedly losing sleep over the implications) framed this as an early warning sign that humans may soon struggle to independently verify AI-generated research output, while skeptics pushed back against hype framing like "physics is solved."
Discussion: 10 tweets from 10 authors · @curiouswavefn, @TricksterKallis, @rad1_study, @WKCosmo, @tmcosmos_org +5