Skip to content

Repository files navigation

FrontierPhysics

Discord GitHub WeChat

Are AI agents good physicists?

FrontierPhysics: Benchmark how AI agents do frontier physics research.

Contributing · Benchmark Protocol · BenchFlow SDK · Discord

What is FrontierPhysics?

FrontierPhysics is a benchmark evaluating how AI agents do frontier physics research iteratively. We evaluate realistic research challenges with iteration loops from literature deep review to research plan implementation. Tasks come from real research problems that take at least weeks of effort for a physics PhD to do deep research and implement, and SOTA LLM agents struggle with. The tasks are evaluated with verifiable graders and per-task rubric-based reviewer agents to make sure agents are doing research in ways aligned with real frontier researchers.

Quick Start

git clone https://github.com/benchflow-ai/FrontierPhysics.git
cd FrontierPhysics

# Install or upgrade to the latest stable BenchFlow CLI.
uv tool install --upgrade benchflow

# Install repository tooling from the committed lockfile.
uv sync --locked

# Validate a native task.md package.
bench tasks check tasks/surface-ion-trap-shuttling

# Oracle must pass before agent runs.
bench eval run \
  --tasks-dir tasks/surface-ion-trap-shuttling \
  --agent oracle \
  --sandbox docker

Runnable benchmark tasks live under tasks/. FrontierPhysics uses uv.lock for reproducible repository tooling while the bench CLI runs task validation and evaluations.

See experiments/README.md for paired no-skill and with-skill commands.

API Keys

Running hosted agents may require provider credentials or an authenticated local agent session. Export only the credentials required by the selected agent. Keep secrets in an ignored .env or .envrc; never commit them.

Creating Tasks

FrontierPhysics tasks are native BenchFlow task.md packages:

tasks/<task-id>/
  task.md
  environment/
    Dockerfile
    skills/
  oracle/
    solve.sh
  verifier/
    rubric.json
    test.sh
    test_outputs.py

See CONTRIBUTING.md for scientific-quality requirements, mentor-skill policy, metadata, validation, and review evidence.

Get Involved

Contribution credit is tracked in points: a merged task you authored is worth 6, referring a contributor is worth 2 once their first task merges, a merged task you reviewed is worth 1, and 12 points earns co-authorship on the FrontierPhysics paper and dataset. Reviewing and referring count — see the authorship policy.

License

Apache 2.0. Bundled third-party components retain their own license notices; see NOTICE.

About

FrontierPhysics: Benchmark how AI agents do frontier physics research.

Topics

Resources

Contributing

Stars

4 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages