ICML 2026 in Seoul: a conference at capacity

The cap

ICML ran from July 6 to 11 at COEX in Seoul, with tutorials on the Monday, the main conference Tuesday to Thursday and workshops on the Friday and Saturday. The thing everyone talked about in the hallways was announced six weeks earlier. On May 24 the organisers published a registration update saying they had set a cap marginally below the venue's legal limit, after weighing space, WiFi and logistics, and that registration was running faster than in previous years.

The mechanics of the cap matter for who got in. Registrations were held back for authors of accepted papers who had not yet registered, for sponsors with complimentary passes, and for workshop organisers. General tickets were described as very limited and expected to sell out. Full refunds were available until June 15 and none after. Virtual registration had no cap, and the notice was careful to say that in-person attendance was optional for publication, though every paper without an attending author needed at least one virtual registration.

Kamalika Chaudhuri, the ICML board president, wrote that the conference has always been built around the quality of scientific exchange and that setting a cap was not a decision taken lightly. We believe her. We also think the cap is the honest version of something that was already true. Halls with a legal occupancy limit have been the binding constraint on these meetings for a few years, and the only change is that it is now written down.

A winter of reviewing trouble

The capacity notice was the second unusual communication from ICML this cycle. The first came on December 15, a statement to reviewers on community standards and support, issued after what the organisers called the OpenReview exploit attack in late November. It condemned harassment, threatening behaviour and misuse of information, and said that collusion, bribery, misrepresentation and fake identities would not be tolerated. Three days later ICML announced it was doubling its annual contribution to OpenReview from 100,000 to 200,000 dollars, on the grounds that the attacks had shown how critical that infrastructure has become. The blog does not describe what the exploit was, and we are not going to guess.

Then there was the LLM policy. ICML 2026 let authors pick one of two reviewing regimes for each paper. Under Policy A reviewers could not use language models at all. Under Policy B they could use privacy-compliant models to understand a paper and polish prose, but could not ask a model to judge quality, list weaknesses or draft the review. Reviewers declared a preference, the matching system honoured it where it could, and anyone who submitted a Policy A paper had to review under Policy A themselves.

On March 18 the organisers reported how enforcement went. They embedded hidden instructions in submission PDFs, drawn from random phrase pairs out of a 170,000 word dictionary, that a language model would follow more than 80 percent of the time. Every flagged review was checked by a person. The result was 795 reviews, about 1 percent of the total, from 506 reviewers who had agreed not to use models and did. All of those reviews were removed. Because 398 of those reviewers had their own papers in the system, 497 submissions were desk-rejected, and 51 reviewers who had used a model on more than half their assignments were dropped from the pool entirely.

What NeurIPS is trying instead

NeurIPS has chosen a different answer to the same size problem. The 2026 meeting runs as a main conference in Sydney from December 6 to 12, with satellite events in Atlanta and Paris from December 9 to 13. The workshop numbers give a sense of how it is being distributed. 102 workshops were accepted from 454 valid proposals, with acceptance rates of 21.5 percent in Sydney, 25.4 percent in Paris and 26 percent in Atlanta. Sixteen competitions were selected. Position papers must now be substantially human-written, with AI limited to copy-editing.

The NeurIPS site does not say why three cities, and we will not put words in the organisers' mouths. But the two designs are legible as two answers to one constraint. ICML kept one room and rationed the seats. NeurIPS added rooms and accepted that the community will be in three places at once. Neither is obviously right, and the fact that the two largest meetings in the field made opposite choices in the same year tells you the format is under real strain.

Who the cap actually hits

Authors were protected. Sponsors were protected. The people who were not are the ones who need the hallway most. A first-year PhD student without a paper yet, a master's student deciding whether to apply, a researcher from a small lab who came to find collaborators rather than present. Being in the room is how you learn which problems people actually care about, which is different from which problems get accepted, and it is how you get introduced to the person who will read your next draft.

Virtual registration is the offered substitute and it is not the same thing. The talks are recorded anyway. What the cap removes is the poster session conversation and the dinner, and those do not stream. If ICML stays capped, and we think it will, the practical advice for early-career people changes. Get a workshop paper, since workshops had separate capacity and were not full in May. Register the week registration opens. Treat the three-city NeurIPS as three smaller conferences and pick the one where your subfield will be.

What we would like the organisers to try

Two things. First, publish the numbers. The capacity notice gives no figure for the cap or for how many people were turned away, and the March post gives counts for violations but not for total reviews. It is hard to argue about whether the format has outgrown its venues without knowing by how much. Second, reserve a slice of general registration for people without accepted papers who are within their first two years of research, selected by lottery. Authors already have a guaranteed seat. The cap should not quietly turn the conference into an authors-only event.

The honest reading of this year is that the reviewing system and the venue system both hit limits at the same time, and both organisations responded with policies rather than with new infrastructure. The policies are reasonable. They are also stopgaps. A field that produces enough papers to need a 170,000 word watermark dictionary to police its reviewers is going to need something other than one convention centre and one December.