On November 4, 2026, a team of 72 authors posted FrontierPhysics to arXiv, an end-to-end benchmark of agents doing physics research. It contains 56 tasks written by 42 physicists from their own research, spanning nine domains across theoretical, computational and experimental physics. Agents must review literature, plan research, run experiments and produce structured outputs and manuscripts. Grading combines deterministic verifiers with expert-written rubrics.
The authors evaluated 22 models in 26 configurations. No configuration passed more than 46 of its 168 rollouts. On manual review, the authors conclude that frontier agents’ main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.
Why it matters: Earlier physics benchmarks largely test question answering. Grading the whole research loop, from literature to manuscript, measures something closer to what labs mean when they talk about AI scientists, and shows how far that still is from routine success.
What it does not show: 56 tasks is a small set, and rubric grading of manuscripts carries judgment. The results are a snapshot of the models and harnesses tested in early November 2026.