Writing analysis code from the papers — filling a gap in particle physics preservation
Collider experiments supply Rivet routines as part of preserving their measurements, yet only 39% of measurements have them documented and public. This work builds an automated workflow on large language models to supply the missing ones.
Paper overview (our summary)
- Field (arXiv category)hep-ex（+2）
- AuthorsAntonio J. Costa, Caterina Doglioni, Christian Gütschow, et al. (5)
- Submitted2026-06-11
- arXiv ID2606.13535v1
Key points
- AgentRivet is an automated workflow extracting physics analysis information from published papers and writing the missing Rivet routines.
- Only 39% of collider measurements have documented and publicly available Rivet routines.
- The workflow includes intermediate code and physics reviews as autonomous quality control.
- Results are reported with commercial large language models from OpenAI, Anthropic and Google on two recent ATLAS and CMS measurements.
- The majority of physics-implementation issues arose from subtle but ambiguous definitions in the publications themselves.
1The gap was counted first
This work starts from a figure: 39%. Collider experiments supply Rivet routines to preserve their measurements, but only 39% of measurements have one documented and publicly available. For the rest, no published means exists of setting the measurement against theory.
2What Rivet carries
- 1A measurement is madeAn experiment publishes a model-independent measurement
- 2A routine is writtenA Rivet routine reproducing that measurement is prepared
- 3Theory is comparedNew theoretical models are set against the measurement through that routine
- 4The tools improveMonte Carlo event generators are developed and tuned, and searches beyond the Standard Model proceed
Without the routine, a measurement cannot be set against theory. Published as a paper it may be, but with no procedure for comparison available, it is hard for later work to use. The 39% is a statement about availability in that sense.
3An automated workflow
What marks the design is that quality control sits inside the workflow. Generation is not the end of it; correctness as code and correctness as physics are reviewed along the way by the workflow itself.
4Where the difficulty lay
The routines produced are reported to carry few syntax errors, with reasonable physics fidelity following the explanations in the publications. Implementation issues arose all the same, and the authors separate out where they came from.
Most of the trouble lay in how the papers were written. Where a definition is subtle but ambiguous, neither a person nor a model arrives at the same implementation. Attempting the automation made visible where the information meant to be preserved is going missing.
Why it matters
Preserving a measurement means leaving behind not the numbers but a procedure by which theory can later be set against them. That such a procedure exists for only 39% of measurements shows publication and reusability to be different things. Finding that most implementation trouble traced back to ambiguous definitions in the papers is a result about the precision of writing mattering more than the limits of the tool.
FAQ
What is Rivet?
Why does the 39% matter?
What caused the implementation issues?
Sources (primary)
Source: arXiv (descriptive metadata is CC0 public domain). Summaries are our own; see arXiv for the original text and PDF.
- arXiv abstract page (original, official)
- PDF (arXiv)
- arXiv ID: 2606.13535