Research resources · public release

Docs & data

Use the public catalog, source repository, and citation resources to inspect the benchmark without crossing the evaluator’s private trust boundary.

Public result catalog

Machine-readable index of published tasks, runs, and result bundles.

Open catalogGitHub

Source repository

Task environments, benchmark tooling, documentation, and license.

View repositoryBibTeX

Citation file

A stable citation for the public PhySciBench benchmark release.

Download citation

Public evidence contract

What a result bundle contains

Published

Task and run identifiers
Official J and reporting-only S
Sanitized probe trajectory and final status
Approved visual artifact and provenance digests

Withheld

Private evaluator implementation and raw internals
Credentials, environment values, and infrastructure details
Agent transcripts and private scientific artifacts
Unverified intermediate outputs

Release status

PICBench public snapshot · September 2026

26 released tasks: 16 design and 10 implementation
25 published design runs from one model–agent pair
10 archive-verified single-task implementation results published
Design J_acc reporting calibration remains provisional

Cite

PhySciBench

@misc{momeni2026physcibench,
  author = {Ali Momeni},
  title = {PhySciBench: A Physical-Science Design Benchmark},
  year = {2026},
  url = {https://physcibench.com}
}

Reuse

License and attribution

PICBench is released under the MIT License. Cite the benchmark when using its task definitions, evaluation framework, published results, or derived figures. A research paper has not yet been linked from this public release; the repository and versioned site remain the authoritative public references.