FrontierPhysics: Evaluating Agents for End-to-End Physics Research (Nov 2026)

Title: FrontierPhysics: Evaluating Agents for End-to-End Physics Research (Nov 2026) Link: http://arxiv.org/abs/2611.06910v1 Date: November 2026 Summary: Introduces FRONTIERPHYSICS, a benchmark of 56 expert-designed tasks across nine physics domains for evaluating AI agents on end-to-end research. Tasks require literature review, research planning, experiments or computation, structured outputs, and a scientific manuscript; grading combines deterministic verifiers with expert-authored rubrics. Across 22 models and 26 configurations, the best reported pass rate is 27.4%. Manual analysis identifies weak rigor, expert intuition, literature synthesis, and instruction following as common failure modes. Key Topics: AI agents for scientific research Physics research benchmarks End-to-end research workflows Deterministic verification and rubric-based grading Agent evaluation and failure-mode analysis Scientific reasoning, rigor, and expert intuition Chapters: 00:00 - Introducing Frontier Physics 01:37 - Why Older Benchmarks Fall Short 02:52 - Designing Open-Ended Tasks 04:04 - End-To-End Research Outputs 04:41 - Grading Scientific Work 05:25 - Combining Verifiers And Review 08:36 - Validating The Rubrics 09:39 - Model Performance And Costs 11:23 - Analyzing Failure Modes 12:24 - Rigor And Self-Verification 13:19 - Missing Expert Intuition 15:36 - Literature And Grading Errors 17:42 - Benchmark Maintenance Challenges 18:24 - Findings And Research Limits 19:28 - Who Defines The Questions? Footage credits: https://kaboompics.com/: https://www.pexels.com/video/women-ta... AI-generated narration. Footage marked illustrative is context, not evidence for the source's claims.