NeurIPS Started Desk-Rejecting Papers Over References That Don't Exist

1. What the thread actually said, and what we didn't take on faith

The claim, as it was circulating: NeurIPS 2026 ran an automated reference checker across submissions, authors received a report visible on OpenReview, and at least one area chair described a "2 or more hallucinated references" bar for automatic desk rejection. A named account describing themselves as an area chair said they'd "had to reject 2 papers in my batch that would definitively had gone through if the authors did not let their LLM access their bibtex." Other commenters reported appeal emails, per-citation appeal processes, and confusion about whether the whole batch of decisions was even out yet.

None of that names a tool, a vendor, or a methodology, and secondhand quotes from a discussion thread are exactly the kind of claim this publication exists to check rather than repeat. So before writing anything, we went looking for the primary sources: the conferences' own statements, the academic literature on the underlying problem, and at least one specific fabricated citation we could verify with our own hands.

2. It's real, and it's bigger than one thread

It is. Four major venues this cycle — NeurIPS, ICLR, ICML, and ACL — have all treated a hallucinated citation as grounds for desk rejection, and three of the four have published enough about their process to check^[1][2][3].

ICLR's program chairs wrote the most detailed public account of any venue we found. Their system "automatically extracted references from a given submission and checked them against multiple standard bibliographic databases and a standard web search." Flagged references then went through three separate rounds of human review — area chairs first, then the program chairs themselves — before a paper was actually desk rejected, "together with an appeal channel to take care of any erroneously flagged papers." They also stated plainly that the process "partially explains the relatively high desk rejection rate at this year's ICLR compared to past years"^[1].

ACL's program chairs published a statement confirming that more than 100 accepted papers were desk-rejected over citations to works that do not exist, using "an automated system... to identify potential citation issues," with every flag then reviewed by program chairs and senior area chairs to confirm the reference really doesn't exist before a paper was rejected^[2]. ICML's Peer Review Ethics page lists "hallucinated references" as grounds for desk rejection of the individual submission^[3]. NeurIPS's own Main Track Handbook is thinner than the other three — a single sentence acknowledging that "there have been many cases of hallucinated citations in literature review, which violates the NeurIPS Code of Conduct," pointing to a separate academic integrity policy we could not access in full^[4].

What NeurIPS's public documentation doesn't spell out, a NeurIPS 2026 area chair did, in public, under their own name. Danish Pruthi posted that in his own area chair batch, "5 out of 8 submissions have 2+ hallucinated references and will likely be desk rejected," adding that citations "are easy to verify and hallucinations can be caught" — and, in a follow-up, that the bar is real but not absolute: "NeurIPS is quite liberal in how they define hallucinations. One of the remaining three papers has a fabricated citation, but that's not enough for desk-rejection"^[5]. That is close enough to the Reddit thread's "2 or more" claim that we're treating it as the real, if informally documented, NeurIPS threshold this cycle.

3. How big is "hallucinated," actually

The most rigorous number we found comes from an independent large-scale audit, not from any conference itself: 48,095 accepted papers and 2.6 million extracted references across ICLR, ICML, NeurIPS and USENIX Security from 2021 through 2026, checked against Semantic Scholar, OpenAlex, Crossref, DBLP and the ACL Anthology, with suspicious matches escalated to a web-search-equipped model before being counted^[6]. Their definition is deliberately narrow — "non-existent works and substantial author-list mismatches" only, explicitly excluding ordinary bibliographic drift like a changed venue name or a year off by one, the kind of thing StrictCite's own engine treats as a much lower-severity discrepancy than an outright fabrication.

Read the right-hand column against this cycle's "2 or more" desk-reject bar: that is roughly one in twenty accepted NeurIPS papers from last year that would have qualified, had this policy existed then. It didn't. This is the first cycle it does, on papers written in the year rates were already, per the same study, at their highest point yet.

None of this started this year, either. A separate study published in February looked at 100 fabricated NeurIPS 2025 citations in detail and found every single one used more than one deception strategy at once — a wrong author list dressed up with a real-looking venue, an invented DOI paired with a title close enough to something genuine to survive a skim^[8]. What changed this cycle isn't the fabrication. It's that four venues finally started checking for it before publication instead of after.

The same audit is also where the "peer review doesn't catch this" claim actually gets checked, rather than assumed. Comparing reviewer scores on affected versus clean papers on the same 1–10 or 1–5 scale each venue uses, the gap was +0.02 at ICLR, +0.005 at ICML, +0.04 at NeurIPS — noise, not signal. Acceptance tier showed no pattern either: posters, spotlights and orals were affected at nearly identical rates. Three to five expert reviewers per paper, and a fabricated citation moves the outcome not at all.

4. Three named examples, and one we checked ourselves

GPTZero, an AI-text-detection company, ran its own hallucination-detection tool against 300 of the roughly 20,000 ICLR 2026 submissions and published named examples^[7]. In a paper it called "MixtureVitae," rated 8.0 by reviewers, a citation's first three authors matched a real paper — the remaining seven names on that citation did not appear on the real paper at all, and some did not appear to correspond to any real person. In a paper titled "TamperTok," also rated 8.0, one citation's entire author list was wrong for a paper that does genuinely exist. And in a submission on AI-assisted medical triage, one reference read, in full: "[3] K. Arnold, J. Smith, and A. Doe" — an invented author list attached to an invented journal.

We didn't take GPTZero's word for that last one. We queried Crossref's public API directly for that exact author combination:

That is a small, single check, and it proves nothing about the other 49 papers in GPTZero's sample or the hundreds more they expect to find. What it does establish is narrower and, we think, worth exactly as much as it is: one specific, published, named claim about a specific fabricated citation, independently re-checked against a different registry than the one used to originally flag it, came back the same way. That is the standard every claim in this piece is trying to meet, including the ones about other people's checkers.

5. The part that should slow everyone down

Every venue that published real detail about its process also admitted the same weakness. ICLR's program chairs, in their own retrospective: their automated system "had a significant false positive rate — for example, it would flag a reference to a paper with a non-English title that a submission's authors had translated to English"^[1]. The independent 48,000-paper audit put it more bluntly: of the candidate hallucinations its authors checked by hand, "the majority were set aside as recoverable references," most often because of PDF-extraction noise mangling an author string, a very recently posted paper not yet indexed anywhere, or a citation to genuinely unpublished work^[6]. Their own conclusion: "an individual flag should never be treated as conclusive without review."

ACL's outcome is the clearest evidence that this tension is not hypothetical. More than 100 papers were desk-rejected over hallucinated references — and then, per an update posted to the same statement, "it has been decided in a meeting with the ACL executive that affected papers must go through the normal resubmission process" rather than face the original decision as final^[2]. A policy adopted, enforced at scale, and then walked back mid-cycle is what it looks like when an institution is building the plane while flying it. The Reddit thread's confusion about appeal timing and whether decisions were even final reads less like impatience and more like an accurate read of the actual situation.

None of this is an argument against checking citations. It's an argument that checking them is exactly the kind of claim that needs a receipt, every time, before it costs someone a paper — which is the same standard we hold our own engine to, and the reason every verdict StrictCite returns names the specific rule, the specific registries, and the specific field-level disagreement behind it rather than a bare pass or fail.

6. What this actually changes

Before this cycle, a hallucinated co-author or an invented citation was, at worst, an embarrassment a careful reader might eventually notice. As of 2026, at four major venues, it is a desk rejection — decided by an automated first pass, at rates the venues' own numbers say catch real problems roughly one time in four to five papers, with a false-positive rate every venue that measured it called significant, adjudicated under appeal processes that are visibly still being figured out in public. Checking your own bibliography before you submit stopped being a nice-to-have the day that combination became true. That's the whole reason a tool like this exists: not to replace the three humans ICLR now has reading every flagged reference, but to be the fourth reader, before them, who tells you what they're going to find.

Check a bibliography for free at strictcite.com, or check a single suspicious co-author before it becomes someone else's finding: the Fake Co-Author Checker.