Publications Automation

The pipeline harvests publication metadata, regenerates the publication pages and adds news bullets to the About page.

Who owns which file

Precedence when the same work appears twice: overrides > manual bib > auto bib. Records are matched by DOI, then arXiv id (eprint), then title. When you paste a published version by hand, always add eprint = "<arXiv id>" so it merges with the preprint record instead of showing up twice.

Sections and classification

/publications/ has three sections, in this order: Preprints, Publications, Workshop Papers.

The decision lives in exactly one place, markdown_generator/publications_model.py, and is written into the type: field of each generated file. The Liquid template only groups by that field – it must never re-derive anything from venue strings.

Rules, applied in order:

1. an entry in _data/publication_overrides.yml wins, always;
2. a workshop-looking venue is a workshop paper unless the record carries a publisher DOI – that is what keeps a Springer volume or journal special issue of a workshop under Publications;
3. an arXiv-only record is a preprint;
4. everything else – journals, book chapters and main-track conferences (ICML, NeurIPS, ICLR, MSML, AISTATS, COLT, L4DC, UAI) – is published.

“Symposium” is deliberately not a workshop marker. Extra venue rules can be added in _data/publication_sources.yml (main_venue_markers, main_venue_acronyms, workshop_patterns) without touching code.

Fixing a wrong entry

Wrong section or ugly venue name: add an override, keyed by arXiv id, doi: or title:.

overrides:
  "2505.07311":
    type: published
    venue: "Scientific Machine Learning: Emerging Topics (SEMA SIMAI Springer Series)"

Missing conference or workshop version: paste the BibTeX into markdown_generator/publications_manual.bib and add the eprint field.

Never edit _publications/*.md or markdown_generator/publications.bib by hand – both are regenerated.

Sources

• arXiv author page: https://arxiv.org/a/muller_j_3.html plus the arXiv API (discovery)
• DBLP title search (ML conference and workshop metadata)
• Crossref (journal DOIs and venue metadata)
• ORCID 0000-0001-8729-0466 and zbMATH ai:muller.johannes.3 are configured for future use

Discovery is arXiv-driven, so anything that never went to arXiv, and any conference version that DBLP does not return for a title search, has to come from publications_manual.bib. Adding your DBLP author id (pid) to _data/publication_sources.yml would let the bot enumerate the conference records directly – that is the single biggest recall improvement still open.

Local run

No third-party dependencies beyond requests and PyYAML (the pybtex dependency was removed; BibTeX parsing lives in markdown_generator/bibio.py).

python3 -m pip install requests pyyaml
python3 markdown_generator/test_publications.py      # offline, no network
python3 markdown_generator/update_publications.py    # network: arXiv, DBLP, Crossref
python3 markdown_generator/pubsFromBib.py
python3 markdown_generator/update_about_news.py

CI workflow

.github/workflows/publications-pipeline.yml, every Monday (17 6 * * 1 UTC) and on manual dispatch. It runs the tests before and after the update and opens a PR on branch bot/publications-pipeline.

Tests

markdown_generator/test_publications.py covers the BibTeX parser, the classification rules (including the workshop-special-issue regression), record merging, and consistency of the checked-in bibliographies: every work appears exactly once and no non-preprint shows an arXiv venue.

What you need to do

• Review the weekly PR.
• Correct anything wrong via publication_overrides.yml or publications_manual.bib, never in the generated files.
• Add your DBLP pid when convenient.