HEPML-LivingReview: A Frozen Bibliography of Machine Learning for Particle Physics

The gap this fills is citation discovery, not code

Machine learning papers in particle physics appear across arXiv, journal sites, and conference proceedings, and a researcher entering a subfield such as jet tagging or calorimeter simulation has no obvious single place to start. The Living Review addresses that by collecting references into what the README calls "a nearly comprehensive list of citations" for people developing or applying these methods in experimental, phenomenological, or theoretical analyses. The audience is narrow and specific: physicists who need to know what has already been done, and who want it grouped rather than dumped into a search result. The README is explicit that the project chooses "a minimal number of categories" so the list stays usable, and notes that a paper can appear under more than one category. That design decision matters. A broad taxonomy would fragment the list into dozens of near-empty headings, so the maintainers traded granularity for navigability. The README also carries a disclaimer worth quoting in spirit: listing a paper does not endorse or validate it, and that judgement belongs to the community and peer review. That sentence is the honest framing of what this document is. It is an index compiled by people, not a peer-reviewed assessment, and the classification reflects the state of the field at the moment the review was frozen.

How the LaTeX source and INSPIRE API fit together

The repository is TeX, which tells you most of the architecture before you open a file. The review is a LaTeX document whose body is a structured list of entries, each entry a citation with a link to arXiv or a DOI. The README credits the INSPIRE REST API as part of how the review was built, alongside the HEP-ML community and named moderators: Benjamin Nachman, Matthew Feickert, Claudius Krause, John Andrew Raine, and Ramon Winterhalder. The presence of the INSPIRE API in that list suggests literature records were pulled from INSPIRE rather than typed by hand, though the README does not describe the build pipeline in detail, so the exact degree of automation is not something I can confirm from this material. What is visible is the output side: a compiled PDF hosted under the GitHub Pages site at iml-wg.github.io/HEPML-LivingReview, a per-year publication plot stored at docs/assets/per_year.png, and a BibTeX file, HEPML.bib, that carries the citation key for the archived version. The workflow implied by that layout is a maintainer adding entries to the TeX source, regenerating the plot, and rebuilding the PDF for the Pages site. The review is grouped by topic rather than by chronology, and the README states that papers may be referenced in more than one category, which means the source cannot be a simple one-entry-one-section file. Any automated processing of the TeX has to tolerate duplicates across sections.

Getting the PDF and citing the frozen snapshot

There is nothing to install. You either read the rendered document or you clone the source. The README offers a direct download link to the compiled review at https://iml-wg.github.io/HEPML-LivingReview/assets/hepml_review.pdf, and the repository itself is at github.com/iml-wg/HEPML-LivingReview. If you want the source, clone the default branch, which is master, and build the TeX with your usual LaTeX toolchain; the README does not prescribe a specific engine or a Makefile target, so treat the build as standard LaTeX until you inspect the source. For citation, the README gives the instruction directly: use \cite{krause_2026_21626667} in HEPML.bib for the final archived version. That is the whole integration story. There is no package to add, no configuration key to set, no service to run. The one operational detail that matters is which artifact you point readers at. The README labels the Zenodo deposit as the final archival version and the release list shows v2026.06.01 as the "Final Archival Snapshot" dated 2026-07-27. If you cite the live GitHub Pages PDF instead, you are citing a document the project itself has stopped updating, and the freeze means the two will not diverge further, but the archival DOI is the stable identifier the maintainers chose.

The freeze is the limitation, and it is deliberate

The README opens with an important notice stating that the project is no longer actively maintained and that literature coverage is frozen as of 1 June 2026. That single fact governs every use case. If you are starting a project in late 2026 or later, this review cannot tell you what the current state of the art is, because the state of the art moved on after the freeze. The per-year publication plot in docs/assets/per_year.png will show a curve that simply stops. There is no mechanism described for accepting community pull requests to extend coverage past that date, and the archival release name makes the intent clear. A second limitation is subtler. The README warns that inclusion is not endorsement, which means a reader who treats the list as a quality filter is misusing it. Highly cited work and marginal work sit in the same section. Third, the topic grouping is a snapshot of how the maintainers saw the field at freeze time; if a subfield reorganized itself afterward, the headings will not reflect that. None of these are defects in execution. They are the natural consequences of ending a living document, and the maintainers handled the ending cleanly by archiving rather than leaving a half-updated repository.

Against arXiv listing pages and INSPIRE search

The obvious alternative is to skip the review and query INSPIRE directly, or to follow the arXiv listings for hep-ph and hep-ex. The difference in approach is real. INSPIRE search is live, filterable, and complete up to the present, and it is the same data source the review drew on. What it does not give you is a curated grouping chosen by people who work in the field, or a single PDF you can hand to a new student. The review's value was always the editorial layer on top of the database, and that layer is now dated. A second comparison is to the conventional reviews the README points to, such as the Snowmass 2021 Computational Frontier report or the 2018 Deep Learning and its Application to LHC Physics review. Those are written arguments with prose and synthesis. The Living Review is a list. If you need explanation, the reviews are the right artifact; if you need a citation set, the list is faster. The honest position is that INSPIRE has now absorbed most of the practical function, since it is live and the review is not. The review's remaining edge is that someone already decided which papers belong together under a heading like jet substructure or unfolding, and that decision is preserved in the frozen PDF.

What the licence does not tell you

The repository metadata supplied here does not include a licence identifier. That is not the same as having no licence, but it does mean you cannot assume terms. For a bibliography this matters in a specific way: the citations themselves are facts and are not the thing under copyright, but the compiled document, the category structure, and the prose in the README are authored material. If you want to reuse the review's section headings in your own document, or redistribute the PDF, the absence of a stated licence is a question you have to resolve with the repository rather than assume an answer. The Zenodo deposit may carry its own terms, which the README does not restate here. I am not in a position to give legal advice, and the material does not settle the question. Treat licence verification as a step before any redistribution, and check the repository and the Zenodo record directly rather than inferring from the topics list. The TeX source being visible in the repository does not by itself grant reuse rights.

Maintenance cost after the freeze is zero, and that is the point

For a project that is no longer maintained, the upgrade cost is the interesting number, and here it is effectively nothing. There are no dependencies to bump, no API keys to rotate, no CI to keep green. The INSPIRE API integration described in the README was a build-time concern for the maintainers, not a runtime concern for users. The final release, v2026.06.01, is labelled an archival snapshot, and the Zenodo deposit carries a DOI, which means the artifact is designed to sit unchanged. If you vendor the PDF into a course page or an internal wiki, it will not break. The cost you carry is the opposite of maintenance: it is the cost of the review going stale in your own materials. A reading list built on this document will need a supplement for anything after 1 June 2026, and that supplement is your responsibility, not the project's. The maintainers chose to close the project rather than let it decay quietly, and the release naming and the Zenodo deposit both reflect that choice. For a bibliography, that is a reasonable way to end.

Editorial conclusion

Adopt it if you need a single citation list for ML-in-HEP work published before 1 June 2026, and cite the Zenodo snapshot with the BibTeX key krause_2026_21626667 rather than pointing readers at the GitHub PDF. Do not adopt it as a current-awareness feed: the README states the project is no longer actively maintained and coverage stops at the freeze date, so anything from mid-2026 onward is simply absent. Before you rely on it, check the topic headings in the source against the subfield you care about, confirm the licence situation with the repository itself, and treat the category assignment as a starting point rather than a settled classification.

Official sources

1. iml-wg/HEPML-LivingReview on GitHub
2. Issues
3. Project website
4. README
5. Releases