What we find checking real bibliographies.
Not case studies we set up to make a point. Real papers, real reference lists, run through the same thirteen registries the workspace uses — and written up honestly when something turns up, registry error or citation error alike.
ICLR 2027 Submission Checklist: Page Limits, LaTeX Templates, AI Disclosures, and Desk-Rejection Traps
The definitive author guide and verification checklist based on official ICLR 2027 author guidelines, LaTeX packages, and AI policies. Covers the 9-page limit, template integrity, reciprocal reviewing, and pre-rebuttal citation audits.
ICLR Desk-Rejects Over Bad References. Its Own LaTeX Template Has Shipped One Since 2018.
Every ICLR 2027 formatting rule and desk-rejection ground, quoted from ICLR's own guidelines and template. The guidelines give two different page limits for the submission version, and the template's sample bibliography has listed Yoshua Bengio twice on a three-author book for nine cycles.
NeurIPS 2026 Decisions Are Out. Pangram’s Own Report Says the Version That Screened the Position Papers Called Up to 14.9% of AI-Polished Human Reviews “AI.”
Twelve widely repeated claims about the Pangram desk rejections, checked against primary sources, including four we got wrong ourselves. What Pangram's own benchmark of v3.3.2 shows, and what an ICML 2026 paper says about the rejections by name.
NeurIPS Desk-Rejected 178 Position Papers for Being “AI-Generated.” The Detector Flagged the Track Chairs Too.
18.4% of all submissions to the track were desk-rejected on the basis of Pangram scores. When one researcher tested the track chairs' own recent papers, the detector scored them at 24–69%. The thresholds, the calibration, and what it means for authors. Corrected 26 September.
NeurIPS Started Desk-Rejecting Papers Over References That Don't Exist
A thread of secondhand rejection stories turned out to trace back to something real: NeurIPS, ICLR, ICML and ACL all made hallucinated citations grounds for desk rejection this cycle. We traced the claims to the conferences' own statements, an independent 48,000-paper audit, and one fabricated citation we re-checked against Crossref ourselves.
What AI Reference Checkers Actually Catch (and What They Don't)
Twenty-two test citations, five real citation checkers, and then the same twenty-two run against our own engine — including two things it got wrong, one closed and one still partial. A DOI that resolves to the wrong real paper turned out to be the hardest thing in the whole test to catch.
A Federal Tax Court Brief Cited Four Cases. Three Don't Exist.
A federal judge found that three of the four cases in a taxpayer's brief were AI-hallucinated, each one landing on a real citation that belongs to somebody else's case. The same failure mode we've documented in academic bibliographies, now showing up in Tax Court filings.
Zotero's BibTeX Export Silently Rewrote a Citation's Title
A real, unedited Zotero library put a functioning HTML tag's attributes into a paper's own title as literal text, doubly escaped an ampersand, and recorded a file's byte count as its page count. What each fault actually is, and how a title that arrives already broken has to be read.
Attention Is All You Need Has Five Fake Twins on Crossref
A citation-spam operation deposited the most-cited deep learning paper in the world five separate times under a fabricated date. What we found checking it, and ten other real papers' real bibliographies.