Agents & tool use

FrontierPhysics

by Bingran You, Wenjun Ke, Ria Rosenauer et al. (BenchFlow, UC Berkeley and 35 other institutions)

What is FrontierPhysics? FrontierPhysics is an agents & tool use eval made by Bingran You, Wenjun Ke, Ria Rosenauer et al. (BenchFlow, UC Berkeley and 35 other institutions). An end-to-end physics research benchmark of 56 tasks built by 42 physicists from their own projects across nine domains. Agents get a PI-style brief, then must search the literature, plan, run simulations or data analysis, and deliver both checkable numbers and a paper manuscript. Grading combines deterministic verifiers with expert-written rubrics (blocking and scored criteria) applied by a reviewer agent. In the v0.1 paper, 22 models in 26 configurations were run three times per task; the best, Claude Opus 5.5 in Claude Code, passed 46 of 168 rollouts (27.4% pass@1), and failures traced mostly to weak rigour and expert intuition rather than missing physics knowledge.

Page last updated 2026-10-18 · Evals Review

Editor note

An end-to-end physics research benchmark of 56 tasks built by 42 physicists from their own projects across nine domains. Agents get a PI-style brief, then must search the literature, plan, run simulations or data analysis, and deliver both checkable numbers and a paper manuscript. Grading combines deterministic verifiers with expert-written rubrics (blocking and scored criteria) applied by a reviewer agent. In the v0.1 paper, 22 models in 26 configurations were run three times per task; the best, Claude Opus 5.5 in Claude Code, passed 46 of 168 rollouts (27.4% pass@1), and failures traced mostly to weak rigour and expert intuition rather than missing physics knowledge.

What it tests
Whether research agents can take an open physics question to a correct, publishable result with a manuscript.
Original source
arxiv.org/abs/2611.06910
Last updated
arXiv v1 2026-11-04
License
Not checked yet — see source

Facts come only from the linked source (checked 15–16 Oct 2026). Credit belongs to the creators. We don't score or rank evals: how we review.

Reviews

No reviews yet.

Reviews are written by people who sign in with GitHub, one per account per eval. We can't confirm that each reviewer used the eval; we remove reviews that break our rules, never just for being negative, and we never add, buy or edit reviews. Averages, when shown, come only from these reviews.

Sign in with GitHub to review

Related agents & tool use evals

  • GAIA — General assistant skills: browsing, file handling, multi-step reasoning.
  • WebArena — Autonomous web navigation and task completion.
  • OSWorld — Computer-use agents operating real operating systems and apps.
  • Terminal-Bench — Agents working in the shell end to end.

All Agents & tool use evals · Comparison table

Discussion

No one has discussed FrontierPhysics yet. Have you run it? Share what it caught, what it missed or how you set it up.

Sign in with GitHub to comment

Be specific and kind. No promotion or affiliate links. Community rules · All discussions