1. Where ICLR 2027 is right now
Abstracts were due on 18 September and full papers on 25 September. The guidelines are explicit about what happens after that: “Submissions can be edited on OpenReview until the submission deadline, following which edits are not allowed till the release of the reviews.” The FAQ adds that “no edits to the paper can be made after the submission deadline, no matter how small”[1].
Reviews are released on 5 November, publicly. From 5 to 18 November, authors can revise the paper. If they do, “a pdfdiff will be applied to compare new changes to the paper against the original submission,” and “area chairs and reviewers reserve the right to ignore changes that are significantly different from the original paper.” Decisions come on 16 December, after which all submissions, accepted, rejected or withdrawn, are de-anonymised and the reviews are made public[1]. A paper withdrawn after the deadline stays on OpenReview in a public “withdrawn papers” section and is “de-anonymized immediately”[1].
In practice: the PDF you uploaded on 25 September is the one that gets checked. You cannot change it now. You can find out what is in it.
2. The formatting rules, in ICLR’s words
Each row quotes the source linked in the same row. [1] is the author guidelines, [2] is the official LaTeX template (the .tex instructions and the .sty file), [3] is the AI policy for authors.
| Rule | What ICLR says | Source |
|---|---|---|
| Page limit, submission | “At the time of submission, the main text should be 9 pages or fewer.” “This limit will be strictly enforced. Papers with main text beyond the page limit will be desk-rejected.” (But see section 4.) | [1] |
| Page limit, revision and camera-ready | “During the discussion/rebuttal phase and for the camera ready, the page limit will be increased to 10 pages.” | [1] |
| References | “The list of references does not count towards the page limit, and unlimited additional pages are allowed for the bibliography/references.” | [1] |
| Appendix | “As many pages of appendices (after the bibliography) as they wish, but reviewers are not required to read the appendix.” | [1] |
| Template | “Authors are required to use the ICLR LaTeX style files … Please make sure you use the current files and not previous versions. Tweaking the style files may be grounds for rejection.” | [2] |
| Text block and margins | “The text must be confined within a rectangle 5.5 inches (33 picas) wide and 9 inches (54 picas) long. The left margin is 1.5 inch (9 picas).” “All pages should start at 1 inch (6 picas) from the top of the page.” The style file sets the paper to 8.5 × 11 inches (US Letter). | [2] |
| Font | “Use 10 point type with a vertical spacing of 11 points. Times New Roman is the preferred typeface throughout.” Title: “17 point, in small caps and left-aligned.” First-level headings: “small caps, flush left and in point size 12.” | [2] |
| Abstract | “Indented 1/2 inch (3 picas) on both left and right-hand margins … The abstract must be limited to one paragraph.” | [2] |
| Anonymity | “Any paper where author identity is revealed in either the main text or the supplementary material will be desk rejected.” Your own arXiv papers “should be cited in third person.” | [1] |
| Review-mode header | In review mode the style file prints “Anonymous authors / Paper under double-blind review” and the header “Under review as a conference paper at ICLR 2027”. The \iclrfinalcopy switch prints your names. The template’s own comment on it: “Uncomment for camera-ready version, but NOT for submission.” | [2] |
| Code and demo links | Must be “completely anonymous. Also, make sure that the host website does not track visitors.” | [1] |
| AI use statement | Required. “A mandatory section in their paper (this section will not be counted towards the page limit).” The template adds: “This statement should not be more than 1 page.” | [3] [2] |
| Ethics statement | Recommended, “at the end of the main text before references”. It “will not count toward the page limit, but should not be more than 1 page.” | [1] |
| Reproducibility statement | Recommended, “paragraph-long”, and it “will not count toward the page limit.” | [1] |
| Acknowledgements | “Is the acknowledgement section included in the page limit? No.” | [1] |
| Citing other ICLR 2027 submissions | “OpenReview provides BibTeX entries that do not list the authors, but does give the title, year and url.” | [1] |
3. Desk-rejection grounds that are not in the PDF
Several of the stated desk-rejection grounds have nothing to do with the paper itself. All of these are quoted from the author guidelines[1]:
- Authors: “no authors can be added or removed after the abstract submission deadline.”
- OpenReview profiles: every author needs one before the full-paper deadline, and “Incorrect information on your profile will be grounds for desk rejection.”
- Reciprocal reviewing: “If none of the authors are registered as a reviewer, it will result in desk rejection for the paper.” Authors on three or more papers must review at least six, and those who “fail to produce complete, high-quality reviews by the rebuttal stage may have their paper submissions desk rejected.”
- Quotas: no more than 20 papers per author, and at most one paper per author where no author is an eligible reciprocal reviewer. Over quota, “we will randomly desk-reject excess submissions until every author is under quota.”
- Abstracts: “placeholder or duplicate abstracts will be removed.”
- Sanctions: “Submissions discovered to be affiliated with entities on this list will be desk-rejected.”
4. Two places the guidelines contradict themselves or say nothing
The page limit for the submission version. The guidelines’ formatting section says: “At the time of submission, the main text should be 9 pages or fewer.” The template says the same: “a strict upper limit of 9 pages for the main text of the initial submission”[1][2]. Further down the same guidelines page, the camera-ready section says: “Page limit is identical with the submission version (10 pages).” And the FAQ, answering how many pages a rebuttal revision may have: “The page limit is identical with the submission version (10 pages)”[1]. The submission version cannot be both 9 and 10 pages. The two later sentences look like text carried over from an earlier year and never updated. We cannot confirm that, so we just record that the page says both. The 9-page rule is the one with “desk-rejected” written next to it, and it is the one the template repeats. Plan around 9.
Whether references will be checked again. We searched the full text of the ICLR 2027 author guidelines, AI policy for authors and call for papers on 4 October. None of them contains the word “hallucinated”, and none describes an automated reference check[1][3]. The closest rule is in the AI policy: “a substantial falsehood, instance of plagiarism, or misrepresentation produced by an LLM would be considered a Code of Ethics violation on the part of the paper’s authors, and might lead to desk rejection of the paper.” The same policy puts “format references” on its list of AI uses where disclosure is recommended rather than required[3]. So the 2027 rules do not say whether last cycle’s reference check will run again. They do say a fabricated reference is your responsibility.
5. What ICLR did about references last cycle
The ICLR 2026 program chairs described their process in a retrospective on 31 March. Their system “automatically extracted references from a given submission and checked them against multiple standard bibliographic databases and a standard web search.” Area chairs did a first round of human review of flagged references, then the program chairs checked every flagged reference themselves: “each paper with flagged references was reviewed by at least three humans.” Then: “all papers with confirmed hallucinated references were desk rejected, together with an appeal channel to take care of any erroneously flagged papers”[4].
They were also open about the system’s weakness: it “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).” And the process “partially explains the relatively high desk rejection rate at this year’s ICLR compared to past years”[4].
What counts as “hallucinated” matters here. The largest independent audit we know of, covering ICLR, ICML, NeurIPS and USENIX Security, defines it narrowly as “non-existent works and substantial author-list mismatches,” and excludes ordinary bibliographic differences[9]. Keep that definition in mind for the next section.
6. The reference in ICLR’s own template
The official ICLR 2027 style-file package ships a sample bibliography, iclr2027_conference.bib, with three entries. The template’s example text cites all three, so they print in its References section whenever the template is compiled as shipped[2]. This is the third entry, exactly as shipped:
Deep Learning has three authors: Ian Goodfellow, Yoshua Bengio and Aaron Courville. The BibTeX entry the authors themselves recommend lists those three, and has no volume field[7]. The template lists Bengio a second time, as fourth author, and adds a “volume 1”.
How long it has been there. ICLR keeps its templates in a public GitHub repository. The duplicated author was added on 2 July 2018, when the ICLR 2019 template was prepared[5]. We checked the sample bibliography for every year in that repository: 2019, 2020, 2021, 2022, 2023, 2024, 2025, 2026 and 2027. The line is identical in all nine[6]. The template has been updated for eight new conference years since then, and the entry has not been corrected.
It reaches published papers. Rendered by the ICLR style, the entry reads: “Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning, volume 1. MIT Press, 2016.” That exact reference appears in the bibliography of a paper published at ICLR 2023[8]. We cannot tell whether its authors took the entry from the template or from somewhere else, and it is not their error to answer for. It is one example. We did not try to count how many papers carry it.
What our own checker says. We ran the template’s three-entry bibliography through StrictCite’s engine on 4 October. One entry verified cleanly. Two were flagged, and one of those two flags is trivial:
| Template entry | What the record says | Our engine |
|---|---|---|
| Hinton, Osindero & Teh (2006), A Fast Learning Algorithm for Deep Belief Nets | Title, authors, year, journal, volume and pages all match Crossref[11]. | Verified |
| Bengio & LeCun (2007), Scaling Learning Algorithms Towards AI | Crossref gives the title as “Scaling Learning Algorithms toward AI”[10]. One letter. We would not call this an error. | Flagged, trivial |
| Goodfellow et al. (2016), Deep learning | Three authors, not four[7]. Our engine compared the list with OpenAlex and reported that it differs. | Flagged, real |
What this is, and what it is not. It is not a hallucination. The book exists, every author named is a real author of it, and under the audit’s definition a repeated name is probably not a “substantial” mismatch[9]. We do not know how ICLR’s own checker would have treated it, and we are not suggesting anyone would be desk-rejected for it. What it shows is narrower. An author list that does not match the record has sat in the one file ICLR tells every author to start from, through nine conference cycles, without being fixed. Last cycle, ICLR desk-rejected every paper with a confirmed fabricated reference, after review by at least three humans[4]. The template it hands to every author shows how easily such errors go unnoticed, and why it is worth checking a bibliography mechanically rather than trusting that someone would have noticed.
7. What you can still do before 5 November
Check the bibliography you submitted, now. You cannot change the PDF until 5 November[1]. If something in it is wrong, it is better to know before a reviewer or a program chair tells you. If it is a real fabrication, you will know what you are dealing with. If it is a false positive of the kind ICLR described, you will have the evidence ready. ICLR 2026 had an appeal channel for wrongly flagged papers[4]. The 2027 guidelines do not describe one.
Know which entries a checker will flag. Two kinds are predictable. Citations to other ICLR 2027 submissions use OpenReview’s BibTeX, which by design lists no authors[1]. And non-English titles translated into English were the false-positive example ICLR itself gave[4]. If you cite a paper in translation, keep the original title in the entry as well.
Disclose if an LLM touched your references. “Format references” is on ICLR’s list of uses where disclosure is recommended[3]. The AI use statement is mandatory anyway, so say it there.
When you revise, stay inside the rules. The revision limit is 10 pages, and the pdfdiff will show reviewers exactly what changed[1]. Use the 2027 files, leave the style file alone, and do not switch on \iclrfinalcopy: it prints your names, and a revealed identity is a desk-rejection ground[1][2]. If you built your paper from the template, look at whether the sample entries are still in your .bib and still cited.
8. Where StrictCite fits
StrictCite checks references, and nothing else on this page. It does not count your pages, measure your margins or look for your name. Those you can check yourself in five minutes against the table in section 2. What you cannot easily check by hand is whether every one of fifty or sixty references resolves to a real record with the authors, title and year you gave it. That is what StrictCite does, against Crossref, OpenAlex, DBLP and other registries. The free plan covers 30 references a day on a rolling 24-hour window[12], so a 60-reference bibliography takes two days. Reviews are released on 5 November. Check a bibliography free at strictcite.com.