SCIENCE UNDER AUTOMATION
Collider-Bench
Agents reconstruct LHC analysis pipelines from papers and public tools; average results did not reliably exceed the physicist-in-the-loop solution.
What this result supports
The benchmark tests reconstruction of published analyses, not detector construction or collection of a new collision dataset.
Evidence record: Primary benchmark preprint. Checked: 2026-10-21.
Where this evidence is used
Experimental physics
- Reconstruct the analysis and quantify outputs: Collider-Bench reports uneven reconstruction and a continuing advantage for the physicist-in-the-loop solution.
- Check calibration and systematic alternatives: Alignment and reconstruction cover bounded checks, with expert experimental definitions supplied.
- Establish a new physical interpretation: Reconstructed analysis supports interpretation, while incomplete implementation details and systematic uncertainty still require physicists.