ICML 2026 Recap: What Actually Happened in Seoul

TL;DR: ICML 2026 ran July 6-11, 2026 at the COEX Convention and Exhibition Center in Seoul, not July 5-10 as this page previously said. The 43rd International Conference on Machine Learning pulled a record 23,918 submissions and accepted 6,352 for a 26.6% acceptance rate (checked 2026-08-24, blog.icml.cc). Two papers shared the Outstanding Paper Award, one on diffusion language models and one on log-concave sampling theory. A 2016 DeepMind paper on A3C took the Test of Time Award. In-person registration sold out weeks early.

I track ICML because the papers accepted here show up as shipping features in AI SaaS products 12 to 24 months later. That is the line I care about as a SaaS SEO and AI Search operator, not the trade-show floor. NeurIPS and ICLR sit in the same peer-reviewed academic bracket; NVIDIA GTC does not. The 2026 edition wrapped over a month ago and its 44 workshops have all closed, so this reads as a recap and a pointer to ICML 2027 rather than a “should you register” guide.

The real dates were July 6-11, not July 5-10

ICML’s own conference site puts Expo and Tutorial day on Monday, July 6, the main technical program on July 7-9, and workshops on July 10-11. The “July 5-10” date this page previously carried starts and ends a day earlier than the actual run, which matters if anyone was booking flights from the old copy. Everything downstream (badge pickup, sponsor keynotes, the poster halls) tracks the corrected dates.

A record 23,918 submissions doubled the 2025 pool

ICML 2026 fielded 24,371 papers at the full-paper deadline; after desk rejections and author withdrawals 23,918 stayed in the review pool (checked 2026-08-24, blog.icml.cc + techtimes.com/articles/319243). The 2025 edition capped at 12,107. That is not organic growth; that is the LLM-writing-and-reviewing pressure hitting peer review at full force.

Metric	ICML 2026	Source
Submissions	23,918	ICML awards blog
Accepted papers	6,352 (26.6%)	ICML awards blog
Spotlight papers	536 (2.2%)	ICML awards blog
Oral papers	168 (0.7%)	ICML awards blog
Workshops accepted	44 plus 4 affinity	ICML workshops page

The clean “record year” story hides one uncomfortable footnote. Program chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, and Martin Jaggi disclosed that 497 papers, roughly 2% of all submissions, were desk-rejected mid-review after 398 reciprocal reviewers were found to have violated the conference’s LLM-usage-in-reviewing policy they had explicitly agreed to follow. Every paper those reviewers were responsible for was desk-rejected regardless of its own quality. That is what “record submissions” costs a peer-review system running at this scale, and I would rather report it plainly than round it away.

Six invited talks, one full-circle keynote

Invited speakers span theory, safety, economics, biology, NLP, and societal impact rather than clustering on one theme. Pascale Fung (HKUST, co-founder of AMI Labs) covered conversational and ethical AI. Susan Athey (Stanford GSB, John Bates Clark Medal) brought causal inference into an ML audience. Sham M. Kakade (Harvard Kempner Institute) covered RL and foundation-model training theory; Kakade is himself a past ICML Test of Time recipient, which gave his talk a full-circle quality when a different RL paper won the same award this year.

Aviv Regev (Head of Genentech Research and Early Development) addressed AI/ML integration into drug discovery through Genentech’s “Lab in the Loop” approach. Verena Rieser (Google DeepMind) covered responsible development and alignment of frontier AI models. Arvind Narayanan (Princeton, co-author of AI Snake Oil, TIME 100 in AI list) closed the lineup on societal impact.

The two Outstanding Papers went in opposite directions

ICML does not use the term “Best Paper”; Outstanding Paper Award is the top research distinction. Program chairs picked 53 candidates from reviewer scores and Area Chair nominations, narrowed to a 22-paper shortlist, then handed it to an 11-member committee chaired by Andreas Krause, which settled on two winners and five Honorable Mentions (official ICML 2026 awards announcement).

Zanlin Ni, Gao Huang, and eight co-authors won for “The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models.” They challenge a dominant assumption about diffusion LLMs: that generating tokens in arbitrary order is a pure advantage over left-to-right. On math and coding reasoning tasks, dLLMs exploit that flexibility to skip exactly the high-uncertainty “forking” tokens that matter most, which collapses solution diversity. Their fix, a fixed left-to-right generation order for RL rollouts they call JustGRPO, keeps parallel decoding at inference while avoiding the failure mode.

Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin won for “High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions.” They settle an open question in score-based sampling: whether ε-error can be achieved in polylog(1/ε) steps using only score evaluations, rather than the poly(1/ε) steps older discretization-based samplers required. Their construction (first-order rejection sampling, FORS) delivers an exponential improvement and, as a byproduct, the first polylog(1/ε) gradient-only sampler for general log-concave distributions.

The Outstanding Position Paper Award went to Sarah Ball and Phil Hackemann for “Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit,” arguing alignment methods built to prevent AI harm are dual-use technologies that can be misused for censorship. The Test of Time Award went to Volodymyr Mnih and seven DeepMind co-authors for the 2016 A3C paper (Asynchronous Methods for Deep Reinforcement Learning). The committee’s citation credits A3C’s parallel actor-learner insight as “a major contributing factor to the success of RL in LLM post-training,” which is as direct a line from a decade-old paper to modern LLM training as any Test of Time citation gets.

“Agentic AI” showed up in 60-plus workshop proposals

Workshop chairs Gergely Neu and Courtney Paquette flagged that some variation of “agentic AI” appeared in the titles of 60-plus submitted workshop proposals, a concentration they described as remarkable even by ICML’s volume standards. The final program accepted 44 workshops plus 4 affinity workshops, including a second iteration of “Agents in the Wild” (multi-agent coordination and safety), “Statistical Frameworks for Uncertainty in Agentic Systems” (calibration for agent pipelines), and “AI for Science: AI Scientists, Tools, Co-authors, or Founders?” (what it means for autonomous systems to conduct rather than assist research).

One accepted Oral paper captured the shift concretely. “Do We Need Adam?” found that plain stochastic gradient descent matches or outperforms the AdamW optimizer in the RL phase of LLM training while updating fewer than 0.02% of model parameters, over 1,000 times fewer than AdamW. That is the kind of quiet, expensive-to-verify result that ends up in production tuning code six months later.

The attendance number I could not verify

ICML’s own blog post from May 24 confirmed in-person main-conference and tutorial registration reached capacity weeks before the conference opened, with organizers citing venue and WiFi infrastructure limits at COEX. ICML Board President Kamalika Chaudhuri framed it as making sure attendees “can actually do the things you came to do.” That confirms strong demand, not a headcount.

I could not find an official post-event attendance figure for ICML 2026 Seoul as of 2026-08-24. Submission and acceptance counts (23,918 and 6,352) are auditable proxies for scale, and on that measure ICML 2026 was the largest edition in the conference’s history by a wide margin. I would rather say that plainly than round “sold out registration” into a number I cannot back up.

Was ICML 2026 worth it, and what to do about 2027

For the audience ICML is actually built for (ML researchers, PhD students, applied engineers tracking foundation-model and RL research before it ships), 2026 delivered real substance. Two genuinely different Outstanding Papers plus a Test of Time citation tracing a straight line from 2016 A3C to today’s LLM fine-tuning is the kind of long-arc validation only peer-reviewed academic venues produce.

The honest caveat: unless you are actively publishing or reviewing ML research, the value is diffuse across 6,352 papers and 44 workshops rather than concentrated into one two-hour livestream the way NVIDIA GTC compresses a year of hardware news. And the desk-rejection controversy over reviewer LLM-policy violations is a sign the peer-review system underpinning ICML’s credibility is under real strain at 23,918-submission scale.

ICML 2027’s host city is due to be announced in August 2026. ICML 2028 is confirmed for the Eastern United States. If your work touches RL, LLM training, or foundation-model theory directly, watch the 2026 workshop concentration on agentic AI. The field’s own workshop proposals are telling you where the next two years of research investment are headed. My breakdown of the best MCP servers tracks the agent-infrastructure layer this research tends to feed into a year or two later.