What Happened
For six days in July, the busiest street in machine learning ran through Gangnam. The International Conference on Machine Learning wrapped up its 2026 edition on July 11 at the COEX Convention & Exhibition Center in Seoul, the first time the field's flagship venue has held its main program in Korea. The numbers were the story before anyone gave a talk: the conference received 24,371 paper submissions this year, up from 12,107 in 2025, and accepted 6,352 of them. Attendance passed 10,000, roughly a quarter more than last year, and in-person registration for the tutorials and the main conference sold out weeks ahead of the opening.
The schedule followed the familiar shape: an expo and tutorial day on July 6, the main conference from July 7 to 9, and two days of workshops to close. The International Machine Learning Society ran the event, with Tong Zhang of the University of Illinois as general chair and a program committee led by Alekh Agarwal, Miroslav Dudík, Sharon Li and Martin Jaggi. Keynotes came from Pascale Fung, Susan Athey and Sham Kakade — a lineup that hints at where the field's anxieties currently sit: conversational systems, economics and incentives, and the theory of what these models are actually doing.
The awards, announced on the ICML blog just before the doors opened, gave the week its narrative. Two Outstanding Paper Awards went to work on diffusion models. Five honorable mentions spanned memorization, video generation, deception probes and the mathematics of grokking. A single Outstanding Position Paper Award went to an argument that the alignment research community may be handing censors a set of tools it never meant to build. And the Test of Time Award went back a decade, to a paper most practitioners have used without ever citing.
Why It Matters
A conference is a poor proxy for a scientific field, but it is the only one that publishes its numbers. Submissions have doubled in a year and grown more than twentyfold since 2015. That growth is not a sign of health so much as a sign of pressure — of a discipline where the volume of work now outruns the community's capacity to read it. The acceptance rate has held in the low-to-mid twenties, which means roughly 18,000 papers were turned away this cycle, many of them competent.
The venue matters too, in a way that goes beyond hospitality. COEX sits on Teheran-ro, the corridor that functions as Korea's commercial technology spine, and holding ICML there put ten thousand researchers within walking distance of the country's semiconductor, telecom and platform companies. For the Asia-Pacific research ecosystem, that proximity is the point. Recruiters from Google, Microsoft, Amazon, Meta, Apple and a row of quantitative trading firms worked the exhibition floor, and the talent market they were working is increasingly regional rather than transatlantic.
What the accepted papers cluster around is also worth reading carefully. Diffusion models took both top awards, and not for new architectures — for interrogating assumptions the field had accepted too quickly. The applied tracks leaned into healthcare, the physical sciences and sustainability. And the trustworthiness track, once a side conversation, now carries the questions that determine whether any of this gets deployed at all.
Reaction
The strongest reactions in Seoul were reserved for the two Outstanding Papers, both of which told the field it had been slightly wrong about something it was confident about. "High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions," by Fan Chen, Sinho Chewi, Constantinos Daskalakis and Alexander Rakhlin, settled a long-standing theoretical question: whether a target error can be reached in a number of steps that grows only polylogarithmically, rather than polynomially, using gradient queries alone. Their answer was yes, via a rejection-sampling construction. The practical reading is that the number of denoising steps a diffusion model needs may, in principle, collapse dramatically.
The other winner, "The Flexibility Trap," led by a team including Gao Huang, went after diffusion language models. Arbitrary-order generation has been sold as the headline advantage of dLLMs over autoregressive models. The authors showed that on reasoning tasks the models use that freedom to route around exactly the high-uncertainty tokens where the reasoning actually forks — collapsing the diversity of solutions. Their fix is almost deflating: go back to a fixed left-to-right order for reinforcement learning rollouts, and keep parallel decoding only at inference time.
The award that generated the most hallway argument, though, was the Outstanding Position Paper. Sarah Ball and Phil Hackemann's "The Alignment Community is Unintentionally Building a Censor's Toolkit" makes a claim the room could not comfortably wave away: that the techniques built to steer models away from harm generalize, without modification, into techniques for suppressing speech. The selection committee singled out its refusal to make this about any single country or company. Among the honorable mentions, John Morris and colleagues drew attention with a memorization result that puts a number on an old question — their estimate is that GPT-style models store about 3.6 bits of a distribution per parameter.
What's Next
The Test of Time Award points the clearest arrow forward, which is a strange thing to say about a paper from 2016. "Asynchronous Methods for Deep Reinforcement Learning," by Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver and Koray Kavukcuoglu, introduced the idea that parallel actor-learners could stabilize reinforcement learning without a replay buffer. The award citation makes the connection explicit: asynchronous RL is now a major contributing factor to how reinforcement learning is run during large language model post-training. The infrastructure that trains today's reasoning models descends directly from it.
That lineage explains where the next cycle of work is heading. If RL rollouts are the bottleneck for reasoning models, then questions about which sampling strategy should drive those rollouts — precisely the question "The Flexibility Trap" reopens — stop being academic. Expect the diffusion-versus-autoregressive argument to be settled less by architecture papers than by whichever approach post-trains more cheaply. Expect the safety track to keep absorbing work that would once have been filed under policy. And expect the submission count to rise again, because nothing in the current incentive structure discourages it.
Closing Thoughts
There is something clarifying about a field that gives its highest honor to a paper arguing the community is accidentally building a weapon, and its longevity award to a paper nobody thought of as a safety contribution at all. Both are, in their way, admissions that machine learning research does not get to choose how it is used downstream. The A3C authors were trying to make Atari agents train faster on a CPU. A decade later their idea underwrites the training loop of systems that answer medical questions.
The conference itself is now large enough that no attendee experiences the same event as any other. Ten thousand people, six thousand papers, and a poster hall you could not walk end to end in a day. Whatever coherence the field still has is maintained by exactly these rituals — the awards, the citations, the arguments in the corridor between sessions. Seoul got to host that argument this year, and the argument was mostly about whether the field understands its own tools well enough to keep building them.
한글 요약
머신러닝 분야 최대 학회인 ICML 2026이 7월 6일부터 11일까지 서울 삼성동 코엑스에서 열리고 막을 내렸습니다. 올해 논문 투고 수는 24,371편으로 지난해 12,107편에서 두 배 이상 늘었고, 이 가운데 6,352편이 채택됐습니다. 참가자는 1만 명을 넘어섰으며 본회의 현장 등록은 조기 마감됐습니다. 국제머신러닝학회(IMLS)가 주최했고, 일리노이대 Tong Zhang 교수가 총괄 의장을 맡았습니다.
최우수 논문상은 두 편의 확산 모델(diffusion model) 연구에 돌아갔습니다. 하나는 확산 언어 모델의 '임의 순서 생성'이 오히려 추론의 핵심 분기점을 회피하게 만든다는 점을 지적한 「The Flexibility Trap」이고, 다른 하나는 그래디언트 질의만으로 고정밀 샘플링이 가능함을 이론적으로 규명한 연구입니다. 최우수 입장 논문상은 정렬(alignment) 연구가 의도치 않게 검열 도구를 만들고 있다고 경고한 Sarah Ball·Phil Hackemann의 논문이 받았습니다.
가장 상징적인 수상은 2016년 딥마인드의 「비동기 심층 강화학습(A3C)」에 주어진 시간의 시험상이었습니다. 당시 게임 에이전트를 빠르게 학습시키려던 이 아이디어는 현재 대규모 언어 모델의 사후 학습(post-training) 강화학습 구조의 토대가 됐습니다. 서울이 아시아 AI 연구 지형에서 차지하는 무게를 확인시켜 준 한 주였습니다.
참고: ICML 2026 수상작 발표 · ICML 공식 홈페이지