AI agents are becoming increasingly important in physics research, yet existing benchmarks usually focus on fixed questions, isolated sub‑tasks, or reproducing papers with given solutions, and they rarely assess research quality together with correctness. To address this gap we introduce FrontierPhysics, an end‑to‑end benchmark that contains 56 tasks created by 42 physicists from their own work, spanning nine sub‑domains across theoretical, computational, and experimental physics. Each task requires an agent to review the literature, devise a research plan, carry out research or experiments, and produce structured results along with a full scientific manuscript. Evaluation combines deterministic verifiers with expert‑authored rubrics, checking both correctness and publishability. Across 22 models in 26 configurations we performed 168 rollouts; the best configuration succeeded in fewer than 46 cases. Manual analysis of agent trajectories shows that the main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.
Review: FrontierPhysics offers a systematic platform for measuring AI performance in realistic scientific workflows, highlighting current models' deficiencies in deep reasoning and intuition and pointing to concrete directions for future improvements.