TLDR Tech / News
New Benchmark TheoPhysEval Tests AI in Physics Problem Solving
A new evaluation framework named TheoPhysEval, comprising 50 original theoretical physics problems at the research level, has been developed to gauge the capabilities of AI agents in autonomously solving complex scientific challenges. When tested with readily available agents and practical tools, the top performer, GPT-6-Astra operating within Codex, attained 98% accuracy at a cost below $3 per problem. The study also observed minimal accuracy decline when agents were restricted from using web searches or reference materials, suggesting a high degree of information recall from existing scientific literature.
Read the full story