FULL STORY

ICLR Submissions Top 60K, Straining Peer Review

ICLR submissions surged past 62,000—tripling last year and exceeding all prior years combined—as AI-generated papers swamp peer review, prompting fierce debate and reform proposals including hybrid human-AI reviewing.

2026-09-19 ~ 2026-09-22 · 6 episodes · 67 posts

Episode 1 · ICLR 2026 Submissions Top 50,000 as AI Slop Overwhelms Peer Review (2026-09-19, 13 posts)

With about a week left before the abstract deadline, researchers reported between Sept 19–21 that ICLR 2026 submissions had surpassed 50,000, with later figures of 55,000 and 58,000—an all-time high. Fears that AI-generated low-quality papers (AI slop) are overwhelming peer review have erupted, and remedies such as per-author submission caps and accepting only the top 300 papers annually have entered the debate; no official figure has been released, and the total may be inflated by NeurIPS resubmissions.

Confirmed

  • On Sept 19, @miniapeur reported submissions had exceeded 50,000 and were still growing, lamenting an incoming flood of AI slop and saying he did not even want to review for this cycle (m1 and m12 repeat the same information)
  • On Sept 20, @menhguin cited data of roughly 58,000 submissions—a record—joking that much of it counts as "AI for Science"
  • On Sept 21, MIT researcher Igor Gilitschenski (@igilitschenski) said submissions topped 55,000, using the hashtag #slopGrenade to mock the flood of low-quality AI-generated papers
  • @gklambauer (Greg Klambauer) and @MichaelDMoor both noted the 50,000-level figure, calling it a "DDoS attack" on machine learning research
  • @DimitrisPapail reported submissions nearing 45,000, remarking that AI makes mass-produced submissions easy—like flipping coins to get lucky acceptances—while what the field truly needs are interesting, cool, slightly crazy papers
  • AlexiGlad (via @burnytech) noted the 50,000 figure, predicted up to 100,000 in the future, and dubbed it the era of "AI slop"
  • @qberthet confirmed the submission surge is indeed coming
  • Roei Herzig (via m6, m9) observed that every few years the community declares reviewing broken while submissions double again; with ICLR past 50,000, he proposed capping per-author submissions, arguing no one has more than 5 conference-worthy ideas per cycle—echoed by Hilde Kuehne
  • ShiboChenTech (via m4) proposed that AI conferences simply accept the top 300 papers each year and stop fixating on acceptance rates

Unconfirmed

  • Reported figures range from "nearly 45,000" to "over 50,000," "over 55,000" and "about 58,000"; no official number has been published, so the final scale awaits confirmation
  • @qberthet cautioned that the overlap between NeurIPS decisions and the ICLR abstract deadline means many pending NeurIPS submissions are resubmitted to ICLR, inflating the total and making it a noisy signal of true submission activity

Why it matters

  • Exponential submission growth far outpaces reviewing capacity, and AI-generated content has driven the marginal cost of low-quality submissions toward zero; @miniapeur and others even assert peer review cannot survive in its current form
  • If mass-produced lottery-style submissions become the norm, the value of top-conference acceptance signals and the academic evaluation system will be diluted. Per-author caps (Herzig's 5 per cycle) and more radical reforms like ShiboChenTech's "top 300 per year" proposal are emerging as candidate responses—prompting open frustration from many frontline researchers

Episode 2 · ICLR 2027 Submissions Top 62K, Exceeding All Previous Years Combined (2026-09-19, 35 posts)

ICLR 2027 abstract submissions exceeded 62K at the September 19-20 deadline, roughly tripling last year's 19,525 and surpassing all editions from 2013-2026 combined. Official statistics are pending, but submission numbers and multiple researcher reports point to the same magnitude. The mismatch between reviewer capacity and the paper flood has hit an extreme, making AI-generated low-quality papers, LLM-assisted review, and academic incentives the focus of community-wide reflection.

Confirmed

  • Thirteen hours before the deadline, submission numbers were near 51,000 (InvariantnCauchy), reaching about 60K an hour before cutoff; commentators called it a "DDoS attack on machine learning research," and a NeurIPS reviewer shared workshop reviews lamenting declining quality.
  • Kyoto University's chuangma reported 62K+ submissions at the abstract deadline; trawasthiai, submitting five minutes before cutoff, posted submission number 62263, implying a total under 63K.
  • mariyaivasileva noted registered submissions topped 60K versus 19,525 last year; extrapolating last year's rates (3.99% desk-rejected, 25.82% withdrawn at abstract stage), roughly 42K papers may reach review.
  • Google DeepMind's Denny Zhou confirmed the total exceeds all previous years combined, adding he was not surprised; Google's gspandana called it a true stress test for peer review and LLM-assisted review. Cited posts note ICLR received only 500 submissions a decade ago, a 120x increase.
  • At a 20% acceptance rate, 12,000 papers would be accepted, prompting questions about whether the system still makes sense; pfau suggested ICLR's "mission is complete," and some proposed shutting the conference down. botsandbits criticized the 20-paper per-author cap as absurdly high.
  • François Fleuret declared the conference/peer-review model dead; rajammanabrolu said this may be seen as the beginning of the end of AI conferences. tokenbender said fewer than 60 papers were worth recommending, comparing academia to "Zimbabwe's currency in 2009"; zivravid joked he could not name 60 worth reading, drawing pushback from MIT professor Aviv Tamar.
  • Pangram founder Max Spero questioned how many of the 62K submissions were fully AI-generated. John Wieting made a public bet that ≥90% of final valid papers (over/under 20K) will be AI-generated; adrilemercier earlier estimated nearly 90% may be mostly AI-written.
  • PTenigma estimated reviewing 60K+ papers would take 500,000 person-hours, and reported NeurIPS has banned LLM-written reviews with severe penalties.
  • On causes, bozavlado blamed publication-count incentives rather than AI itself; gleech countered that pre-LLM growth was 40% annually versus 200% this year, a 5x acceleration hard to attribute to incentives alone.

Unconfirmed

  • All figures come from submission numbers and individual reports; official final statistics, effective submission counts, and AI-generation share estimates remain unverified.

Why it matters

  • Submissions exceeding all previous years combined signal a severe disconnect between review capacity and volume, threatening review quality and conference credibility. The debate over AI slop, LLM reviewing bans, incentive structures, and quality dilution has become a collective reckoning for academic ML in the LLM era.

15 more related posts →

Episode 3 · ICLR Submissions Exceed 50,000, Researchers Propose Hybrid Human-LLM Review (2026-09-20, 2 posts)

With over 50,000 ICLR submissions requiring an estimated 150 person-years of review effort, researcher Shayan Kiyani proposes a hybrid human-LLM review pipeline with an AI pre-screening round to address the reviewer shortage.

Episode 4 · ICLR Submissions Surge Past 50,000, Sparking Debate (2026-09-20, 4 posts)

ICLR submissions reportedly reached 50,000–60,000, prompting both celebration of growing research activity and jokes that authors are just inflating each other's Google Scholar citations.

Episode 5 · ICLR 2027 tops 60k submissions, sparking calls for peer-review reform (2026-09-21, 8 posts)

ICLR 2027 received over 60,000 submissions, and researchers took to social media on Sept 21-22 to debate structural reform. Ziv Ravid argued the submission-and-review system must be redesigned from scratch to survive this volume, and MIT professor Phillip Torr publicly declared peer review broken. The debate splits between hard quantity caps and mechanism-based incentives, with some predicting AI will eventually do most reviewing.

Confirmed

  • Submission caps: Ziv Ravid proposed a hard per-author limit (e.g., 3-4 papers), ending the era of submitting 15 papers at once
  • Shorter papers: Return to 3-4 page papers with key results up front
  • Reproducibility by default: Submissions must include code, data and intermediate results, auto-verified by agents; failures are not published
  • Accountable reviewing: Reviewers should bear real monetary costs to improve quality and responsibility
  • Tiered cap proposal: Per AnkaReuel, one scheme would let authors with no top-CS-conference acceptances submit at most 1 paper, those with acceptances up to 3, and PI-level senior authors up to 20
  • Anti-spam governance: sarahookr called for per-author submission limits per conference, sharing organization-level submission statistics to identify where surges come from, and faster isolation of fake accounts before human review, to curb paper mills
  • MIT professor's verdict: Per CSProfKGD's repost, Phillip Torr said ML reviewing has always been random, good reviewers are harder to find as papers surge, and the reciprocal-review norm "is broken"
  • Empirical testing: ymasano proposed vibe-coding a website to test rules like per-author caps or fractional 1/k co-author credit using ICLR's public review data
  • AI reviewing prediction: sytelus replied that Ravid misses the point—most new research will soon be done by AI, so reviewing will likewise be done mostly by AI

Why it matters

60k submissions push the current review system to its limits in both labor and quality; even an MIT professor declares the reciprocal-review norm broken, and paper-mill spam is eroding conference credibility. The central divide is between hard constraints (caps, shorter papers) and incentive mechanisms (reproducibility checks, paid reviewing, fractional author credit). sarahookr targets fake accounts and institutional surges at the source, while sytelus pushes the question further—if AI does the research and reviewing, the reform target may not be the review process itself. Empirical testing on public data could move the debate from positions to evidence.

Episode 6 · ICLR 2027 Sees 60k+ Submissions, Sparking Debate Over Paper Inflation (2026-09-21, 5 posts)

ICLR 2027 received over 60,000 submissions, sparking a sharp debate over whether AI paper volume reflects rampant low-quality output or simply normal researcher productivity.

Confirmed

  • ICLR 2027 abstract submissions exceeded 60,000, a figure cited across multiple posts.
  • Earlence Fernandes estimated that at best only 500 of them contain serious results, arguing that writing papers has become fashionable in AI.
  • Yarin Gal (DeepMind/Oxford), citing kchonyc's estimate of 30,000 AI-focused PhD students worldwide, noted that if half of the 60,000 abstracts are duplicated placeholders due to overlapping NeurIPS decision timing, the effective volume is roughly one submission per PhD student.
  • srchvrs reached a similar conclusion: relative to the researcher population, the submission volume is not anomalous.

Unconfirmed

  • The "half are NeurIPS-overlap placeholders" figure is an assumption with no data on actual duplication rates.
  • The "only 500 serious papers" estimate is Fernandes' personal judgment with subjective criteria; discussants also asked how to distinguish genuine researchers from paper mills, with no answer.

Why it matters

  • The dispute reflects widespread anxiety about quality and review burden amid AI's explosive growth: if most submissions are filler, top-venue filtering mechanisms are under strain; if it's normal per-capita output plus scheduling overlap, the panic may be a misreading. The tension between these two readings is key to understanding the current AI academic ecosystem.