Verify BibTeX Files for Hallucinated References with bib-verify

Where this fits

You’ve just finished a literature review and your agent has generated a references.bib file. Before you drop it into your paper, you need to be sure every entry is real. A single hallucinated citation can sink a submission. bib-verify automates that sanity check by cross-referencing each entry against CrossRef, arXiv, and DBLP, flagging anything that doesn’t match.

Why agents benefit

• Field-level diffing – compares title, authors, year, and DOI independently, so you see exactly which part of a reference is wrong, not just a binary pass/fail.
• Three-source fallback – checks CrossRef, arXiv, and DBLP, increasing the chance of finding the real record even if one database is incomplete or slow.
• Structured output – each entry is labeled verified, suspect, or not_found, making it easy to filter and fix only the problematic ones.
• Zero manual API setup – the skill handles parsing and querying; you only need to provide the .bib file path and optionally a CrossRef email for better rate limits.
• Plugs into existing pipelines – works standalone or as a stage in a larger paper-review workflow, so you can chain it before submission checks.

Practical scenarios

Pre-submission bibliography audit

A research assistant agent compiles references for a grant proposal. Run bib-verify on the final .bib file to catch any fabricated entries before the proposal goes to the review board. The report shows which entries need manual verification and which are clear fakes.

Catching LLM-generated citations in a draft

You used an LLM to draft a related-work section, and it invented a few references that look plausible. Feed the generated .bib file through bib-verify – any entry marked not_found is almost certainly hallucinated, saving you from embarrassing corrections later.

Post-review metadata cleanup

Even real references often have wrong years or missing DOIs. After a journal reviewer points out inconsistencies, run bib-verify on the revised file to see field-level mismatches. Fix only the specific fields that differ, rather than manually checking each entry.

Add it to your agent workflow

Install the skill using the OpenAgentSkill CLI:

npx skills add agentscope-ai/OpenJudge --skill bib-verify

Then, in your agent code, point the skill at a .bib file:

from openjudge import PaperReviewPipeline

pipeline = PaperReviewPipeline(mode="bibtex")
report = pipeline.verify("references.bib", crossref_email="you@example.com")
for entry in report:
    print(entry.key, entry.status, entry.mismatches)

The output gives you a per-entry status and a list of mismatched fields. You can then decide whether to fix the entry or drop it.

Compare before adopting

Before you commit, check how bib-verify stacks up against alternatives:

• Quality signals – Does it use multiple databases? Does it report field-level details or just a pass/fail? bib-verify does both.
• Maintenance freshness – Look at the GitHub repo (linked below) for recent commits and issue resolution. A stale skill may break with API changes.
• Alternatives – There are other citation-checking tools, but many require manual API keys or only check one source. bib-verify is agent-friendly and self-contained.
• Workflow fit – Does it integrate with your existing pipeline? If you already use OpenJudge, the API is consistent. If not, the standalone CLI might be enough.

Why it is worth tracking

bib-verify comes from the OpenJudge repository (809 stars, actively maintained), which signals community trust and ongoing development. As LLM-generated research content grows, the need for automated citation verification will only increase. Track this skill if you build agents that produce or audit academic text – it’s a practical guardrail against one of the most common failure modes. For a broader catalog of agent skills, see the OpenAgentSkill listing. Evaluate it now, and you’ll have a reliable check ready for your next submission.