TheoPhysEval Puts Agents Near Physics Perfection

This is a striking benchmark result, but not evidence that AI has achieved autonomous physics research.

• –Problems span quantum field theory, cosmology, condensed matter, and other advanced areas
• –Astra’s 98% score is impressive, but each agent ran only once
• –Accuracy changed little without web access, suggesting strong memorization of physics literature
• –The authors note that several apparent model errors were mistakes in expert reference answers
• –The harder frontier remains choosing worthwhile problems, recognizing dead ends, and generating new theories