Research Conferences in 2026: How Scale and AI Are Rewriting Peer Review

For most of the last century, the research conference was a supporting act. The journal held the record; the conference was where you met people. In computing and machine learning, that relationship inverted decades ago — the conference is the archival venue — and that inversion is now under measurable stress.

The stress is not rhetorical. Submissions to the largest computing conferences have roughly doubled in a single year at some venues. Reviewer pools have not. Into that gap has arrived a technology the field itself built: large language models capable of drafting a plausible paper, and a plausible review of one, in minutes.

This article is for people whose work depends on that system functioning: researchers deciding where to submit, area chairs and programme chairs designing processes, R&D leaders deciding whether conference travel still returns value, and university and industry staff who fund it. It covers what the evidence actually shows about conference peer review, what the major venues changed in 2025 and 2026, which claims remain contested, and what to do differently as an author, reviewer or attendee. Much of the underlying dynamic — volume growth outpacing volunteer capacity, and prestige concentrating in a handful of venues — is structural and will outlast any particular model or policy.

Executive summary

• Volume growth is the root cause, not AI. NeurIPS submissions grew from 9,467 in 2020 to 21,575 in 2025 (NeurIPS Program Chairs, 2025). ICML 2026 received 23,918 submissions against 12,107 the previous year, accepting 26.6% (RIKEN AIP, 2026).
• Conference review was already noisy before AI. In the NeurIPS 2021 consistency experiment, two independent committees disagreed on accept/reject for 23% of papers, and roughly half the accepted list would change on a rerun — closely matching the 2014 result (Beygelzimer et al., 2023).
• Undisclosed AI reviewing is now measurable at scale. An analysis of all ICLR 2026 reviews estimated 21% were fully AI-generated, with more than half showing some AI involvement (Pangram, 2025); Nature reported the resulting controversy (Naddaf, 2025).
• Venues have moved from norms to enforcement. ICML 2026 desk-rejected 497 papers after 795 reviews (~1% of all reviews) written by reviewers who had agreed to a no-LLM policy were detected via PDF watermarking (ICML Program Chairs, 2026).
• Enforcement now also targets submissions. The NeurIPS 2026 Position Paper Track desk-rejected 178 submissions (18.4%) and required 123 more (12.7%) to produce provenance evidence of human authorship (NeurIPS Position Paper Chairs, 2026).
• Sanctioned AI review is being tested, not adopted. AAAI-26 generated one clearly labelled AI review for all 22,977 main-track papers in under a day at under $1 per paper; surveyed participants rated AI reviews favourably on technical accuracy (Biswas et al., 2026). NeurIPS 2026 is running an opt-in randomised trial instead (NeurIPS, 2026a).
• Geography is being restructured. NeurIPS 2026 runs as one main site in Sydney plus official satellites in Atlanta and Paris, with 102 workshops split across all three (NeurIPS Workshop Chairs, 2026).
• The sustainability and access case for distributed formats is evidence-backed. Virtual conversion reduced conference carbon footprints by about 94%, and well-chosen multi-hub hybrid formats by about 67% (Tao et al., 2021); virtual formats also improved participation across cost, career stage and geography (Skiles et al., 2022).
• What remains unresolved: whether AI assistance improves review accuracy, how to audit AI use without punishing honest authors, and whether prestige can be decoupled from a single annual deadline.

---

Why conferences became the record, not the rehearsal

In most sciences, a conference talk is a preview and the journal article is the citable object. Computing diverged. Fast-moving subfields wanted a fixed annual deadline, full-length reviewed papers, and archival proceedings — so the conference absorbed the journal's function. Acceptance at NeurIPS, ICML, ICLR, CVPR, AAAI or ACL became a hiring signal, a promotion criterion and, in industry labs, a performance metric.

That design has real virtues. It produces a hard, predictable clock; it forces work into public view quickly; and it concentrates a field's attention on a shared set of results once a year. It also produces three structural liabilities that are now binding:

1. A single point of failure. One review cycle, with a fixed deadline, decides a year of career outcomes for thousands of people.
2. Volunteer labour as load-bearing infrastructure. Reviewing is unpaid, largely unrewarded and increasingly required as a condition of submitting.
3. A venue with walls. Unlike a journal, a conference must physically seat the people whose papers it accepts.

Every reform discussed below is an attempt to relieve one of these three.

The scale shock, in numbers

The following table compiles publicly reported figures from official conference sources and institutional summaries. It is compiled for OneWise; it is not reproduced from any single external table. Figures for different venues use different denominators (main track only, or all tracks), so cross-venue comparison should be treated as indicative.

Table 1. Reported submission scale at major computing conferences, 2020–2026

Two observations matter more than any single number. First, growth is compounding, not linear: ICML's 2026 intake was close to double its 2025 intake. Second, acceptance rates have not collapsed. Venues have absorbed volume by accepting more papers in absolute terms — which shifts the bottleneck from selectivity to logistics, and from logistics to review capacity.

Callout — Read acceptance rates carefully. A stable acceptance rate across a doubling of submissions means the accepted-paper count also doubled. Selectivity as a signal is diluted even when the percentage is unchanged. Hiring committees and internal promotion rubrics that treat "NeurIPS paper" as a fixed-value token are, in effect, using a unit whose value has drifted.

What the evidence says about review quality — before AI

It is tempting to treat AI as the cause of unreliable reviewing. The evidence does not support that framing. Conference review was demonstrably noisy when the venues were an order of magnitude smaller.

In 2014, NeurIPS routed 10% of submissions through two independent committees. The experiment was repeated in 2021 at larger scale. The published results: the committees disagreed on accept/reject for 23% of duplicated papers; 50.6% of papers accepted by one committee were rejected by the other (49.5% in 2014); and 14.9% of rejected papers would have been accepted on re-review (17.5% in 2014). The authors concluded that approximately half the accepted list would change on a random rerun, and — importantly — found no evidence that the process had become noisier as the field grew (Beygelzimer et al., 2023; NeurIPS, 2021).

Two implications follow, and professionals frequently draw the wrong one.

• Correct implication: a rejection at a top venue is weak evidence about a paper's quality, and authors should calibrate accordingly. The 2021 analysis itself notes that authors should not be excessively discouraged by rejected work.
• Also correct, and less comfortable: making venues more selective increases arbitrariness. Raising the bar does not filter more accurately; it randomises harder.

This is the baseline against which AI-era reforms should be judged. The relevant question is not "is AI-assisted review perfect?" but "is it better or worse than a stretched human process that already flips half its accept decisions on rerun?" That question is currently unanswered.

The AI inflection: three postures, all in production

Between late 2025 and mid-2026, major venues stopped debating AI in peer review and started running policy at scale. Three distinct postures emerged.

Posture 1 — Detect and enforce. ICML 2026 ran a two-track framework: Policy A (no LLM use) and Policy B (LLMs permitted for comprehension and language polish, but not for identifying strengths and weaknesses or drafting reviews). Reviewers chose, and no reviewer who strongly preferred B was assigned to A. The programme chairs then watermarked submission PDFs with hidden instructions — a method based on peer-reviewed work by Rao et al. (2025) — and detected 795 reviews (~1% of all reviews) from 506 Policy A reviewers that carried the watermark. Where the violating reviewer was a reciprocal reviewer, their submission was rejected: 497 papers, corresponding to 398 reviewers. Every flagged case was manually inspected, and the chairs calculated a family-wise error rate of 0.0001 (ICML Program Chairs, 2026).

The chairs' own caveat is the most useful part of the report: the technique catches only careless use — feeding a PDF to a model and pasting the output. Anyone who paraphrases, drafts partially, or strips the watermark is not caught. The ~1% figure is a floor, not a measurement.

Posture 2 — Audit authorship provenance. The NeurIPS 2026 Position Paper Track required papers to be substantially human-written, with AI limited to copy-editing. Working with a detection vendor under a zero-retention agreement, the chairs found that 28.2% of 969 submissions received a maximum AI score under default settings. They then did something unusual and worth emulating: they stress-tested the detector against permissible and impermissible use cases, and against pre-ChatGPT control papers from another venue, before acting. They also re-ran detection with smaller text windows to reduce over-claiming. The outcome was 178 desk rejections (18.4%) and 123 submissions (12.7%) required to supply a version-history audit trail — a "pre-AI" checkpoint, a "post-AI" checkpoint, and evidence that AI edits introduced no new substantive content (NeurIPS Position Paper Chairs, 2026).

Posture 3 — Sanction, measure, and keep humans deciding. AAAI-26 generated one clearly labelled AI review for every main-track submission — all 22,977 full-review papers in under a day, at under $1 per paper — and instructed senior programme committee members to cross-check AI-surfaced issues rather than defer to them. Surveyed authors and committee members reported finding the AI reviews useful, and preferred them to human reviews on dimensions including technical accuracy and research suggestions (Biswas et al., 2026). NeurIPS 2026 took a more cautious route: an opt-in, IRB-reviewed randomised experiment in which volunteer reviewers are assigned per paper to unassisted review, open-ended LLM interaction, or structured LLM interaction, with zero data retention by model providers (NeurIPS, 2026a).

Table 2. Comparing the three postures

The honest summary: all three are experiments with published outcomes and unpublished long-term effects. Participant preference for AI reviews (AAAI-26) is a finding about perceived usefulness, not about decision accuracy — the pilot's own framing points toward human–AI teaming rather than replacement. Anyone citing it as proof that AI reviews better than humans is over-reading it.

Reforms that have nothing to do with AI

Two 2026 changes deserve attention because they address the structural liabilities directly.

Contribution types. NeurIPS 2026 asked main-track authors to select a contribution type — General, Theory, Use-Inspired, Concept & Feasibility, or Negative Results — with the same review form interpreted differently for each. The two rarest categories are the interesting ones: Concept & Feasibility explicitly covers high-potential ideas that cannot be fully validated in one paper, and Negative Results covers well-executed findings that something does not work (NeurIPS, 2026b). Both are chronically under-published across science, and both are hard to defend under a generic novelty-and-benchmark rubric.

Journal-first routing. The Machine Learning Reproducibility Challenge became an official NeurIPS 2026 track under a model where papers are first accepted by the Transactions on Machine Learning Research journal, then given a lightweight compatibility check for presentation (NeurIPS, 2026c). This decouples the quality decision from the annual deadline — the single most direct answer to the "one point of failure" problem — while retaining the conference as the place where work is presented.

Reciprocal reviewing ties submission rights to reviewing duty, and NeurIPS has previously allowed reviews to be withheld from authors who did not meet their own obligations, with desk rejection for grossly negligent reviewers (NeurIPS Program Chairs, 2025). It expands the reviewer pool by fiat, but it also expands it with people whose primary motivation is compliance — which is plausibly part of why enforcement actions land disproportionately on reciprocal reviewers.

Geography, access and the carbon question

Physical capacity is now a binding constraint, and venues have responded by distributing rather than centralising. NeurIPS 2026 runs its main conference in Sydney (6–12 December) with official satellites in Atlanta (8–13 December) and Paris (9–13 December); 102 workshops were accepted from 454 valid proposals and split 48 / 28 / 26 across the three cities (NeurIPS, 2026d; NeurIPS Workshop Chairs, 2026). The Paris site, developed on the ELLIS-backed European model, lets authors present accepted papers without registering for the main site.

Three separate arguments support this shift, and they should not be conflated:

• Capacity. Distributed sites relieve a hard physical ceiling on attendance.
• Access. Visa friction is a persistent and unevenly distributed barrier. A Nature comment describes organisers relocating a global conference specifically to reduce visa discrimination against researchers from lower- and middle-income countries (Nature, 2025), and reporting through 2025–2026 documents conferences relocating or cancelling over travel and immigration uncertainty. Regional hubs partially route around this; they do not solve it.
• Emissions. The peer-reviewed evidence here is reasonably strong. Conference travel can reach the order of three tonnes of CO₂-equivalent per participant (Klöwer et al., 2020). Converting to fully virtual formats cut modelled footprints by about 94%, and hybrid formats with carefully selected regional hubs by about 67% (Tao et al., 2021). Virtual formats also measurably widened participation across cost, career stage, gender and geography (Skiles et al., 2022).

The limitation worth stating plainly: a multi-hub model reduces average travel distance, but three simultaneous sites also mean three venues, three sets of local logistics and a fragmented hallway track. The emissions benefit depends on attendees actually choosing the nearest hub. No published evaluation of the three-site NeurIPS model exists yet, because it has not happened yet.

Frequently misunderstood, and three myths

• Myth: "Rejection means the work is weak." The consistency experiments say otherwise — roughly half of accepted papers would not be accepted on a rerun (Beygelzimer et al., 2023). Rejection is a noisy signal, especially near the threshold.
• Myth: "Higher selectivity means higher quality." The same analysis suggests greater selectivity increases arbitrariness rather than precision.
• Myth: "AI detection settles the question." The venues doing the detecting are the most explicit about its limits: ICML's chairs called their method easy to circumvent, and the NeurIPS position track ran control experiments precisely because detector output alone was not considered sufficient grounds for action.

Latest developments (dated)

• 22 September 2025 — Rao, Kumar, Lakkaraju and Shah publish the watermark-based detection framework in PLOS ONE that ICML would later operationalise.
• November–December 2025 — Analysis of all ICLR 2026 reviews estimates 21% fully AI-generated; Nature reports the controversy (Pangram, 2025; Naddaf, 2025).
• 12 January 2026 — ICML announces a self-ranking policy: authors with multiple submissions voluntarily rank their own papers, with author–reviewer disagreement used to flag papers needing attention.
• 14–26 January 2026 — ICML runs Google Research's Paper Assistant Tool for ~4,500 papers pre-deadline. Of 869 survey respondents, 92.1% would use it again; 35.4% of theory-paper authors said it found gaps taking over an hour to fix; 31% ran new experiments as a result (ICML, 2026b).
• 18 March 2026 — ICML publishes its enforcement results: 497 desk rejections (ICML Program Chairs, 2026).
• 16 April 2026 — NeurIPS announces main-track contribution types (NeurIPS, 2026b).
• 15 April 2026 — The AAAI-26 AI review pilot report is posted (Biswas et al., 2026).
• 4 May 2026 — MLRC becomes an official NeurIPS track via TMLR-first routing (NeurIPS, 2026c).
• 2 June 2026 — NeurIPS Position Paper Track announces 178 desk rejections and a provenance appeals process with a 15 June deadline (NeurIPS Position Paper Chairs, 2026).
• 6–11 July 2026 — ICML 2026 convenes in Seoul with 23,918 submissions and a 26.6% acceptance rate.
• 10 August 2026 — NeurIPS announces 102 accepted workshops across Sydney, Paris and Atlanta, with new caps of eight organisers per workshop and two proposals per person. One accepted Atlanta workshop is devoted to authorship, peer review and conference governance under AI — a sign the field is now formally studying its own institutions (NeurIPS Workshop Chairs, 2026).

Practical takeaways

If you submit papers

1. Keep a provenance trail. Draft in a system with version history. If you use AI beyond light copy-editing, be able to show a pre-AI checkpoint, a post-AI checkpoint and your subsequent verification. This is already an appeals requirement at one track and is a plausible default within a few cycles.
2. Read the track policy, not the venue policy. In 2026, tracks within the same conference carried materially different AI rules. Assuming a single conference-wide policy is now an error with desk-rejection consequences.
3. Verify every citation manually. Hallucinated references have been found in accepted papers at top venues. Automated citation checking is not part of the standard reviewer workflow; it must be part of yours.
4. Match the contribution type to the actual contribution. If your result is a well-executed negative finding or an early-stage concept, submitting it as a general paper invites rejection on criteria it was never designed to meet.

If you review or chair

5. Treat AI output as a checklist, not a verdict. The AAAI-26 design instructed committee members to cross-check AI-surfaced issues against human reviews rather than defer to them. That is the defensible posture.
6. If you agree to a policy, honour it. The 2026 enforcement actions attached real consequences — including loss of your own submission — to violating a policy you selected.
7. Publish your process data. The venues that published detection thresholds, control experiments and error rates made their decisions contestable and therefore more legitimate.

If you fund or manage research staff

8. Stop treating one accepted paper as a fixed-value unit. Accepted-paper counts have grown with submission volume. Evaluate contribution, not venue tokens.
9. Budget for hubs, not headquarters. With multi-site conferences, the cheapest and lowest-emission option is often the nearest satellite — and for many purposes it delivers the same programme.
10. Fund reviewing time explicitly. Reviewing is now a compliance obligation attached to submitting. If it is unbudgeted, it will be done badly or delegated to a model.

Key insights

1. Volume growth, not AI, is the primary driver of conference strain; AI is an accelerant and a symptom.
2. Conference review was already noisy at one-tenth the current scale, with ~50% of accepted papers flipping on rerun.
3. Increasing selectivity increases arbitrariness rather than precision.
4. Stable acceptance rates alongside doubling submissions dilute the signalling value of a single accepted paper.
5. Undisclosed AI reviewing has been measured at roughly one in five reviews at one major venue; detection-based enforcement figures are floors, not true rates.
6. Enforcement has moved from norms to consequences: desk rejections in the hundreds at both ICML and a NeurIPS track in 2026.
7. Sanctioned AI review is now deployed at conference scale, but its effect on decision accuracy remains unmeasured.
8. Provenance auditing — version history showing pre- and post-AI checkpoints — is emerging as the practical alternative to detector-only judgement.
9. Structural reforms that decouple quality decisions from the annual deadline, such as journal-first routing, address the root problem more directly than detection does.
10. Distributed multi-hub conferences have a solid evidence base for emissions and access benefits, but the three-site model has not yet been evaluated in practice.

Figure descriptions for the OneWise design team

Figure 1. The conference review pipeline and its 2026 intervention points Purpose: Show where AI now enters a conference review workflow and where governance controls sit. Layout: Horizontal left-to-right flow, five primary stages in rounded rectangles: (1) Author drafting → (2) Submission and desk screening → (3) Reviewer assignment → (4) Review and discussion → (5) Area chair / senior area chair decision → (6) Publication and presentation. Above the flow, place a lighter parallel band labelled "AI touchpoints" with callout markers dropping onto stages 1, 4 and 5: "pre-submission feedback tools" onto stage 1, "LLM-assisted or LLM-generated reviews" onto stage 4, "AI-generated review as one input" onto stage 5. Below the flow, a second band labelled "Governance controls" with markers rising to stages 2, 3 and 4: "AI-use declaration and provenance audit" to stage 2, "policy-matched reviewer assignment" to stage 3, "watermark detection and manual verification" to stage 4. Use arrow direction left-to-right throughout; use dashed arrows for the AI band and solid arrows for governance. Caption: "AI now touches drafting, reviewing and decision support. Governance controls have been added at screening, assignment and review."

Figure 2. Three institutional responses to review overload Purpose: Compare the trade-offs of detection, provenance auditing and sanctioned AI assistance. Layout: Three vertical columns of equal width, each headed by a label — "Detect and enforce", "Audit provenance", "Sanction and measure". Within each column, four stacked rows with consistent labels: Targets, Mechanism, Strength, Limit. Add a horizontal band beneath all three columns reading "Shared unresolved question: does any of this improve decision accuracy?" with upward arrows from each column meeting it. Caption: "Three postures now running in production at major venues, each addressing a different failure mode."

Frequently asked questions

What is a research conference, and why does it matter more in computing? A research conference is a periodic meeting where peer-reviewed work is presented. In computing and machine learning, conference proceedings are archival and count as primary publications — unlike most other sciences, where journals hold that role.

Are conference papers peer reviewed? Yes, at major venues. Papers typically receive three to five reviews, an area chair meta-review and a senior area chair decision. The rigour is comparable in structure to journal review, but compressed into a single annual cycle with no unlimited revision rounds.

How arbitrary is conference peer review? Measurably arbitrary near the acceptance threshold. NeurIPS ran controlled experiments in 2014 and 2021 in which 10% of submissions were reviewed twice by independent committees; the committees disagreed on 23% of papers in 2021, and about half of accepted papers would change on a rerun.

How many reviews at AI conferences are written by AI? One analysis of all ICLR 2026 reviews estimated 21% were fully AI-generated and more than half showed some AI involvement. Detection-based figures from other venues are lower but are explicitly described by their authors as lower bounds, because the methods catch only careless use.

Can I use an LLM to write my paper? It depends on the venue and the specific track. In 2026, policies within a single conference differed by track — one NeurIPS track required papers to be substantially human-written with AI limited to copy-editing, and desk-rejected submissions that appeared non-compliant. Always read the track's own call for papers.

Can I use an LLM to write a review? Usually not without disclosure, and at some venues not at all. ICML 2026 let reviewers choose between a prohibitive and a permissive policy, then enforced the prohibitive one with detection. Reviewers remain fully responsible for review content under every published policy.

How do conferences detect AI-generated reviews? One published method embeds hidden instructions in the submission PDF that a language model — but not a human reader — would act on, then tests whether the resulting text appears in the review. It offers strong statistical control against false positives but is straightforward to circumvent once known.

What should I do if I'm accused of undisclosed AI use? Venues running provenance-based appeals have asked for a document version history showing the work before AI involvement, immediately after the most substantial AI edits, and as finally submitted, plus an analysis showing AI edits added no new substantive content. Maintaining that history prospectively is far easier than reconstructing it.

Is it still worth attending in person? It depends on your objective. For dissemination, virtual and satellite attendance is close to equivalent. For collaboration formation, hiring and unstructured conversation, in-person attendance still carries advantages that virtual platforms have not fully replicated — a limitation that peer-reviewed evaluations of virtual conferences also identify.

What is a satellite or mirror conference? An official secondary site running concurrently with the main conference, where authors can present accepted papers without travelling to the primary venue. NeurIPS 2026 uses this model with sites in Sydney, Atlanta and Paris.

Do virtual conferences actually reduce emissions? Yes, substantially, according to peer-reviewed modelling: about 94% reduction for fully virtual and about 67% for hybrid formats with well-chosen regional hubs. Travel dominates the footprint; streaming is a minor component.

What is reciprocal reviewing? A policy tying the right to submit to the duty to review. It expands the reviewer pool but recruits reviewers whose primary incentive is compliance rather than interest.

What are contribution types? A submission-time declaration of what kind of contribution a paper makes — for example theory, use-inspired work, early-stage concepts or negative results — used to calibrate how reviewing criteria are interpreted.

What is journal-first or TMLR-first routing? A model in which a paper is accepted by a journal with rolling review, then presented at the conference after a lightweight compatibility check. It decouples the quality decision from a single annual deadline.

Should I be discouraged by a rejection from a top venue? The authors of the NeurIPS consistency analysis explicitly concluded that authors should not be excessively discouraged by rejected work, because outcomes near the threshold are substantially driven by reviewer assignment.

Glossary

• Area Chair (AC) / Senior Area Chair (SAC): Volunteers who oversee a batch of papers, synthesise reviews into a meta-review and make or escalate accept/reject recommendations.
• Camera-ready: The final formatted version of an accepted paper submitted for the proceedings.
• Collusion ring: A group coordinating bids and reviews to secure favourable outcomes for each other's papers; explicitly prohibited under anti-collusion policies.
• Consistency experiment: A controlled study in which a random subset of submissions is reviewed by two independent committees to quantify decision randomness.
• Desk rejection: Rejection without full review, typically for policy violations such as formatting, anonymity breaches or undisclosed AI use.
• Double-blind review: A process in which authors and reviewers are mutually anonymous during evaluation.
• Indirect prompt injection: Instructions embedded in a document that influence a language model processing it; used defensively in review watermarking and offensively to manipulate AI reviewers.
• Meta-review: An area chair's summary of reviews and reasoning toward a recommendation.
• OpenReview: A widely used conference management platform supporting open or semi-open review workflows.
• Reciprocal reviewer: An author required to serve as a reviewer as a condition of their own submission.
• Satellite (mirror) site: An official concurrent conference location where accepted papers may be presented.
• Spotlight / Oral: Presentation tiers above poster status, awarded to a small fraction of accepted papers.
• TMLR (Transactions on Machine Learning Research): A journal with rolling review, now used as an upstream acceptance route for at least one NeurIPS track.
• Watermarking (review detection): Embedding covert, model-visible markers in a submission PDF to test whether a review was machine-generated.

---

References

Academic papers and peer-reviewed research

Beygelzimer, A., Dauphin, Y. N., Liang, P., & Wortman Vaughan, J. (2023). Has the machine learning review process become more arbitrary as the field has grown? The NeurIPS 2021 consistency experiment (arXiv:2306.03262). arXiv. https://doi.org/10.48550/arXiv.2306.03262

Klöwer, M., Hopkins, D., Allen, M., & Higham, J. (2020). An analysis of ways to decarbonize conference travel after COVID-19. Nature, 583(7816), 356–359. https://doi.org/10.1038/d41586-020-02057-2

Rao, V. S., Kumar, A., Lakkaraju, H., & Shah, N. B. (2025). Detecting LLM-generated peer reviews. PLOS ONE, 20(9), e0331871. https://doi.org/10.1371/journal.pone.0331871

Skiles, M., Yang, E., Reshef, O., Robalino Muñoz, D., Cintron, D., Lind, M. L., Rush, A., Perez Calleja, P., Nerenberg, R., Armani, A., Faust, K. M., & Kumar, M. (2022). Conference demographics and footprint changed by virtual platforms. Nature Sustainability, 5(2), 149–156. https://doi.org/10.1038/s41893-021-00823-2

Tao, Y., Steckel, D., Klemeš, J. J., & You, F. (2021). Trend towards virtual and hybrid conferences may be an effective climate change mitigation strategy. Nature Communications, 12. https://www.nature.com/articles/s41467-021-27251-2

Conference reports and technical reports

Biswas, J., Schoepp, S., Vasan, G., Opipari, A., Zhang, A., Hu, Z., Joseph, S., Lease, M., Li, J. J., Stone, P., Wagstaff, K. L., Taylor, M. E., & Jenkins, O. C. (2026). AI-assisted peer review at scale: The AAAI-26 AI review pilot (arXiv:2604.13940). arXiv. https://arxiv.org/abs/2604.13940

Official conference documentation and organiser statements

ICML Program Chairs (Agarwal, A., Dudik, M., Li, S., Jaggi, M.), Shah, N. B., Gorman, K., & Kamath, G. (2026, March 18). On violations of LLM review policies. ICML Blog. https://blog.icml.cc/2026/03/18/on-violations-of-llm-review-policies/

ICML. (2026b, March 30). Retrospective on PAT × ICML 2026 AI Paper Assistant Program. ICML Blog. https://blog.icml.cc/2026/03/30/retrospective-on-pat-x-icml-2026-ai-paper-assistant-program/

NeurIPS. (2021, December 8). The NeurIPS 2021 consistency experiment. NeurIPS Blog. https://blog.neurips.cc/2021/12/08/the-neurips-2021-consistency-experiment/

NeurIPS Program Chairs. (2025, September 30). Reflections on the 2025 review process from the Program Committee Chairs. NeurIPS Blog. https://blog.neurips.cc/2025/09/30/reflections-on-the-2025-review-process-from-the-program-committee-chairs/

NeurIPS. (2026a). NeurIPS 2026 AI-assisted reviewing experiment. https://neurips.cc/Conferences/2026/ai-reviewing-experiment

NeurIPS. (2026b, April 16). A choice of contribution types at NeurIPS 2026. NeurIPS Blog. https://blog.neurips.cc/2026/04/16/a-choice-of-contribution-types-at-neurips-2026/

NeurIPS. (2026c, May 4). MLRC 2026: Reproducibility as an official track at NeurIPS. NeurIPS Blog. https://blog.neurips.cc/2026/05/04/mlrc-2026-reproducibility-as-an-official-track-at-neurips/

NeurIPS. (2026d, April 22). NeurIPS newsletter — April 2026. NeurIPS Blog. https://blog.neurips.cc/2026/04/22/neurips-newsletter-april-2026/

NeurIPS Position Paper Chairs (Lu, A., Lazar, S., Rügamer, D., Hua, S., & Metcalf, K.). (2026, June 2). AI-generated papers in the NeurIPS 2026 Position Paper Track. NeurIPS Blog. https://blog.neurips.cc/2026/06/02/ai-generated-papers-in-the-neurips-2026-position-paper-track/

NeurIPS Workshop Chairs (Zamzmi, G., Ricci, E., Sun, S.-H., Doan, K., Papamarkou, T., & Koniusz, P.). (2026, August 10). Announcing the NeurIPS 2026 workshops. NeurIPS Blog. https://blog.neurips.cc/2026/08/10/announcing-the-neurips-2026-workshops/

NeurIPS. (2026e). Main track handbook 2026. https://neurips.cc/Conferences/2026/MainTrackHandbook

Journalism and industry analysis

Naddaf, M. (2025, December). Major AI conference flooded with peer reviews written fully by AI. Nature. https://www.nature.com/articles/d41586-025-03506-6

Nature. (2025, March 4). We moved a conference halfway around the globe to avoid visa discrimination. Nature (Career column). https://www.nature.com/articles/d41586-025-00614-1

Pangram. (2025, November 18). Pangram predicts 21% of ICLR reviews are AI-generated. https://www.pangram.com/blog/pangram-predicts-21-of-iclr-reviews-are-ai-generated

Institutional summaries

RIKEN Center for Advanced Intelligence Project. (2026, May 24). 45 papers have been accepted at ICML 2026. https://aip.riken.jp/news/icml2026/

Note on sourcing: Submission and acceptance figures are drawn from official conference communications where available. Where a figure was reported only through secondary channels, it is attributed to that channel in the text and treated as indicative rather than authoritative. The Pangram ICLR analysis is a vendor-published study rather than peer-reviewed research; it is reported here alongside Nature's independent coverage of the resulting dispute, and its methodology has not been independently replicated.