venue-ref-guard

Rule-based reference integrity checks for AI conference submissions.

Catch hallucinated, malformed, and suspicious references before they become a desk-reject risk.

Quick start · Venues · Checks · Policy basis

Why this exists

LLMs are now part of literature survey and related-work writing, but they can generate references that look real and point to no real paper. Several AI venues now treat non-existent references as a serious integrity issue.

venue-ref-guard gives authors a local pre-submission check:

1. Export your .bib from Overleaf or your LaTeX project.
2. Optionally include the rendered submission PDF.
3. Run the venue profile for ACL, EMNLP, NAACL, ICLR, ICML, NeurIPS, or CVPR.
4. Resolve every suspected_hallucinated, invalid, and needs_review entry before submitting.

This is not a promise of acceptance and it cannot eliminate all desk-reject risk. It is a practical guardrail that makes citation-integrity problems visible while you can still fix them.

Quick Start

Run from the repository:

python3 -m venue_ref_guard venues
python3 -m venue_ref_guard rules --venue icml
python3 -m venue_ref_guard check --venue icml --bib examples/references.bib --output report.md

Check a paper PDF against the same BibTeX file:

python3 -m venue_ref_guard check \
  --venue neurips \
  --bib references.bib \
  --paper paper.pdf \
  --output venue-ref-report.md \
  --json venue-ref-report.json

Run an offline structural check when network access is unavailable:

python3 -m venue_ref_guard check --venue acl --bib references.bib --offline

Supported Venues

Venue	Key	Rule family	Focus
ACL	acl	ACL/ARR	Real artifacts, ACL citation style, pubcheck-adjacent checks
EMNLP	emnlp	ACL/ARR	ACL/ARR integrity policy plus EMNLP paper-integrity rules
NAACL	naacl	ACL/ARR	ACL/ARR citation and publication-ethics rules
ICLR	iclr	OpenReview	LLM-use responsibility and hallucinated-reference desk-reject risk
ICML	icml	OpenReview	Published AC guidance for problematic references
NeurIPS	neurips	OpenReview	Author responsibility for text, figures, and references
CVPR	cvpr	CVF	Cited-reference pages, arXiv version hygiene, reference integrity

What It Checks

Check	Result
BibTeX syntax and required fields	Finds missing title, author/editor, year, and malformed entries
DOI shape and DOI/title disagreement	Flags bad DOI formats and strong-title matches with conflicting DOIs
Public-source existence	Searches Crossref, OpenAlex, Semantic Scholar, DBLP, and arXiv
Fuzzy title reality check	Flags entries when no public source has a title near the submitted title
Author/year/venue drift	Separates minor repair issues from potentially fabricated metadata
PDF/BibTeX alignment	Checks whether BibTeX titles appear in the rendered reference section when PDF text extraction is available
Venue-specific rule mapping	Reports each issue with the relevant official rule URL

Report Statuses

Status	Meaning	Action
verified	Strong public-source title/author/year match found	Keep it, but still review important citations manually
needs_review	Fuzzy match, artifact-like citation, or lookup uncertainty	Manually confirm and update BibTeX from the authoritative source
offline_unverified	Network source matching was skipped	Re-run online before submission
invalid	BibTeX is structurally incomplete for checking	Fix the BibTeX entry first
suspected_hallucinated	No close public-source title match was found	Treat as a blocker until manually disproven

Example Output

# venue-ref-guard report

- Venue: ICML (`icml`)
- References checked: 42
- Suspected hallucinated: 1
- Invalid BibTeX entries: 0
- Needs manual review: 3
- Verified: 38

## Findings

- BLOCKER `NO_SOURCE_MATCH` `smith2026imaginary`
  No public source found a publication with a title close to this reference.

- WARNING `AUTHOR_LIST_REVIEW` `lee2024benchmark`
  Author names differ from the best public source match.

Exit Codes

Use --fail-on to make the CLI suitable for CI or a pre-submission script:

python3 -m venue_ref_guard check --venue iclr --bib references.bib --fail-on suspected
python3 -m venue_ref_guard check --venue iclr --bib references.bib --fail-on needs-review
python3 -m venue_ref_guard check --venue iclr --bib references.bib --fail-on warning
python3 -m venue_ref_guard check --venue iclr --bib references.bib --fail-on never

Mode	Non-zero exit when
suspected	A reference is suspected hallucinated or invalid
needs-review	A reference is suspected, invalid, or needs manual review
warning	Any warning, error, or blocker exists
never	Never fails from report findings

Official Policy Basis

The built-in profiles use public official sources:

Venue family	Sources
ACL/ARR	ACL hallucinated-reference statement, ACL author guidelines, ACLPUB style guide, aclpubcheck
EMNLP	EMNLP 2026 CFP, double-blind instructions, ACL/ARR policy
ICLR	ICLR 2026 author guide, ICLR response to LLM-generated papers and reviews
ICML	ICML 2026 author instructions, ICML 2026 AC hallucinated-reference guidance
NeurIPS	NeurIPS 2026 main track handbook
CVPR	CVPR 2026 author guidelines

Install Notes

The current implementation has no required third-party runtime dependencies.

Optional PDF extraction:

• Install Poppler so pdftotext is available on PATH, or
• install the optional Python extra after packaging:

pip install "venue-ref-guard[pdf]"

Development

python3 -m unittest discover
python3 -m venue_ref_guard check --venue acl --bib examples/references.bib --offline --fail-on never

Project layout:

venue_ref_guard/
  bibtex.py      BibTeX parser
  venues.py      Official venue profiles and rules
  verifier.py    Reference-level verdicts
  sources.py     Crossref/OpenAlex/Semantic Scholar/DBLP/arXiv clients
  pdf.py         Optional PDF/reference-section alignment
  report.py      Markdown and JSON reports
  cli.py         Command-line interface

Design Principle

venue-ref-guard errs on the side of manual review. A single suspected hallucinated reference is reported as a blocking finding even when a venue escalation policy mentions a higher reporting threshold, because the intended user is an author checking before submission.