AxionLimitBench

A benchmark for extracting experimental exclusion limits from particle-physics papers. Given a paper that bounds an axion, dark-photon or other light-boson coupling, a system must decide whether the paper reports a new measured limit, identify the coupling, and return the excluded-region boundary as a curve in a canonical (mass, coupling) plane. Each answer is graded against the curve the maintainer of the AxionLimits compilation committed for that paper.

• 292 graded papers, 13 coupling types. Papers whose reference cannot grade any extraction (prediction bands, published-only data, superseded versions) are excluded with recorded evidence (data/EXCLUSIONS.md).
• The GitHub repository is the canonical home of the benchmark (references, task card, prediction schema, frozen scorer, baselines). This dataset mirrors it and adds what is too large for git: the full tool transcript of every agent session (results/*/*/<id>.events.jsonl, about 3 GB).
• Headline metric: success rate, the fraction of graded papers whose curve is within 10% (0.041 dex) of the reference; abstentions, wrong coupling types and unconvertible declarations count as misses.
• Paper: AxionLimitBench: can AI curate a dark matter constraint repository? (ML4PS 2026, submitted): 10.5281/zenodo.22880507; the PDF is also paper.pdf in this repository.
• Archived on Zenodo: https://doi.org/10.5281/zenodo.22838233

Leaderboard

system	model	success rate	compared	cond. median [dex]	catastrophic	$/paper
generic agent (Claude Code, 7 tools, task card only)	claude-fable-5	56%	97%	0.02	0.0%	1.48
generic agent	claude-opus-4-8	47%	93%	0.04	0.7%	0.99
AutoAxionLimits pipeline	claude-fable-5	19%	90%	0.16	3.8%	0.88
AutoAxionLimits pipeline	claude-opus-4-8	14%	89%	0.18	3.5%	0.44
generic agent	claude-haiku-4-5	9%	75%	0.38	7.3%	0.28
AutoAxionLimits pipeline	claude-haiku-4-5	5%	77%	0.62	17.6%	0.09

Contents

data/            papers.json (labels), ground_truth/*.txt (reference curves), manifest.json,
                 measured_pool_ids.json, EXCLUSIONS.md, PLANE_AUDIT.md, fetch_pdfs.py
docs/TASK.md     the task card every system sees
schema/          prediction JSON schema
scorer/          frozen scorer (python3 -m scorer.score --snapshots DIR)
baselines/       reference agent driver (Claude Code, headless) and leak audit
results/         per-system, per-paper predictions (<id>.json), scores (metrics.json),
                 leaderboard, failure audits, and for agent runs the full session
                 transcripts (<id>.events.jsonl, stream-json)

PDFs are not redistributed; data/fetch_pdfs.py downloads them from arXiv.

Evaluate your own system

1. Read docs/TASK.md. Give your system the PDF (and, if you like, the paper's own arXiv e-print) and nothing else: no compilations, HEPData, INSPIRE or web search.
2. For each id in data/measured_pool_ids.json, write results/<system>/<run>/<id>.json in the prediction schema, plus a run.json with model, date and any caps.
3. python3 -m scorer.validate --snapshots results/<system>/<run> then python3 -m scorer.score --snapshots results/<system>/<run> --report.

Citation

See CITATION.cff. Reference curves are from cajohare/AxionLimits (MIT, Ciaran O'Hare).