dist/index.html (single-file build) or serve this folder over HTTP.A Systematic Analysis of ICLR 2027 Submissions
When submissions double in a year, do the safeguards of peer review keep pace? Only partly. The growth is mostly new work rather than recycled ICLR rejects, and verbatim copying is essentially absent among the pairs our retrieval can surface. But anonymity is nominal for the % of submissions with a same-title arXiv preprint, and the submissions that full-text comparison confirms as duplicates of published or concurrently reviewed work were still under review at our snapshot.
ICLR 2027 received submissions, times the of ICLR 2026. We took a snapshot of the complete pool on , during the review period, and use only information that is public at that time. Author identities, reviews and decisions are not yet public, so every analysis is designed to be re-run when they are.
Primary areas and topics, recycled rejects, and an "automated-research" title style.
2A calibrated audit funnel: abstract retrieval, LLM adjudication, full-text comparison.
3Double-blind exposure, PDF-level leakage, the organisers' response, and where rejected papers go.
About ICLR 2026 rejects reappear at ICLR 2027, against additional submissions, so ICLR's own rejects explain little of the growth. Papers rejected elsewhere, notably at NeurIPS 2026 shortly before the ICLR deadline, are invisible to us and may contribute more. The topical mix, however, shifted sharply.
ICLR 2026 vs. 2027. Hover a row for details.
Foundation models and LLMs remain the largest area ( submissions, ×). Infrastructure and systems (×) and robotics and planning (×) grow fastest; generative models (×) and graphs (×) grow more slowly than the conference.
Share of titles matching each topic. Pick topics below; dashed lines fell in 2027.
Agent-related titles doubled their share, from % to % ( submissions). Memory, world models, vision-language-action and self-evolving systems grew even faster in relative terms. Some keywords appear almost from nothing: on-policy distillation went from to submissions, LLM agents from to . Multimodal LLMs, parameter-efficient fine-tuning, state-space models and quantisation lost share.
Share of titles containing each marker, ICLR 2024–2027. Thick lines: any strong marker, two or more markers.
Template: a negated claim followed by a methodological noun phrase, as in "X Is Not Y: A Pre-registered Audit of …"
We audit every submission against conference and arXiv abstracts with a funnel that runs from cheap, high-recall signals to expensive, high-precision ones, and calibrate each stage against positive and negative controls.
Each stage is cheaper per item than the next and removes most of the remaining candidates. Hover or tap a stage.
The full-text stage covers grade A/B leads from all sources (titles, arXiv metadata and abstracts), not only the grey zone, so it is not a strict subset of the stage above it. Counts are pairs until the full-text stage and submissions from there on.
No submission reproduces the abstract of a published paper under a new title, even with a relaxed threshold. Re-titled matches are mostly benign: re-titled versions of the authors' own arXiv preprints, resubmissions of papers rejected elsewhere, and pairs of ICLR 2027 submissions with identical abstracts.
pairs have similar meaning but different words ( against published work, within ICLR 2027, against arXiv). The second LLM pass keeps only pairs scored as "substantially the same work" against published papers, plus scored as extensions.
Reference lists, figures and runs of result numbers are hard to disguise by rewriting. All random same-area control pairs are near zero on every signal (maximum reference overlap %), and all submissions compared with their own arXiv versions are classified as the same paper.
One verdict per submission, strongest pair kept: submissions ( of pairs with both PDFs).
Double-blind review, submission-time checks and the decision itself are the safeguards of a conference. We measure each with public data.
submissions have an arXiv preprint with an identical title; of these were posted on or after 2026-09-15, around the submission deadline. ICLR permits preprints, so this is not a violation, but it bounds the anonymity that double-blind review can deliver for a large part of the pool.
grade-A (objectively verifiable) submissions had been desk-rejected as of , both of them blank placeholders. All grade-A submissions that full text confirms as the same paper as a published paper, or as a paper declared under review elsewhere, were still under review. Across the pool, submissions were newly desk-rejected and newly withdrawn since the snapshot.
Estimated share of submissions, stratified sample weighted by inverse inclusion probability, with 95% bootstrap intervals.
The submitted PDFs themselves leak little. Simple rules for hidden text (invisible render mode, white fill, sub-1 pt fonts, off-page text) and injection phrases fire on several percent of PDFs, but on inspection almost all hits are benign: white labels on coloured figures, embedded formula data, and papers about prompt injection. The strict rules, which require that nothing is drawn beneath the hidden text or that off-page text addresses reviewers or LLMs, find no case in the sample. % of submissions include a statement on LLM usage.
Every reference resolved against arXiv, accepted-paper lists, DBLP, printed arXiv identifiers, Crossref and arXiv search. ICLR 2027 sample vs. a control sample of ICLR 2024 submissions processed identically.
The share of submissions with at least one unresolved, non-trivial reference is lower than in the 2024 control. The top unresolved cases in both years are real but hard-to-index items (datasets, software, workshop papers), so the control rate is the false-alarm floor of the method. We find no evidence of an excess of fabricated references in ICLR 2027 submissions.
Share later accepted at ICML or NeurIPS 2026, by mean ICLR 2026 rating.
Of the papers ICLR 2026 rejected, identical titles alone find % later accepted; abstract matching raises this to –%, because many papers are re-titled after rejection. The rate rises from –% for mean ratings ≤ 2.5 to –% for 4.5–5.5. Withdrawn papers fare much worse (–%), and only –% of rejected papers reappear at ICLR 2027. NeurIPS 2026 abstracts are not yet public, so these are lower bounds.
Organisers, who unlike outside researchers have access to author identities and the full PDFs, can run such a pipeline at negligible cost. Everything in this study ran on a single node with open-weight models.
With a two-round LLM funnel, the number of pairs that need human review at ICLR scale is on the order of one hundred.
Every analysis uses only information that any logged-in OpenReview user, or any visitor to arXiv, DBLP and the conference websites, could see during the review period. Submission PDFs were downloaded through the public OpenReview API at its published rate limits.
The author serves as an ICLR 2027 reviewer. We used no information visible only to reviewers, area chairs or program chairs: not the identities or size of the reviewer pool, assignments, bids, reviews, discussions or author identities.
This page and the paper report aggregate statistics only. No submission is identified, and no title, number or excerpt of any flagged submission is shown.
Integrity leads come from automated signals with known error rates and have not been adjudicated. We report aggregates only and do not identify any submission, here or in the downloadable data.
The tables behind the charts on this page. Every number on the page is read from paper/numbers.tex, the macro file generated by src/paper_numbers.py that the paper itself uses, so page and paper cannot disagree.