Physicists Launch FrontierPhysics Benchmark for AI Research Agents

The authors of a new arXiv paper have introduced FrontierPhysics, a benchmark that tests AI agents on full physics research workflows instead of isolated questions. Its 56 tasks were written by 42 physicists from their own research and cover nine areas of theoretical, computational and experimental physics. Agents must review literature, plan a study, run experiments and produce structured results and manuscripts, graded by automated verifiers and expert-written rubrics. Across 22 models in 26 configurations, no configuration passed more than 46 of its 168 rollouts. The authors report that frontier agents mainly fall short on reasoning rigor and expert intuition rather than physics knowledge or execution.