ICML 2026 Recap: What Actually Happened in Seoul

TL;DR: ICML 2026 ran July 6-11, 2026, at the COEX Convention & Exhibition Center in Seoul, South Korea - not July 5-10 as the ACF date field on this page previously showed. The 43rd International Conference on Machine Learning pulled in a record 23,918 submissions, more than double 2025’s total, with 6,352 papers accepted for a 26.6% acceptance rate. Two papers shared the Outstanding Paper Award, one on diffusion language models and one settling a long-standing question in sampling theory, while a 2016 DeepMind reinforcement learning paper took the Test of Time Award. Six invited speakers covered everything from conversational AI to drug discovery, and in-person registration sold out weeks before the conference opened.

I track AI research events the same way I test AI SaaS tools before recommending them: check what actually happened against primary sources, not whatever a directory page guessed months in advance. ICML is not a corporate keynote conference like NVIDIA GTC - it is one of the “big three” peer-reviewed academic venues in machine learning, alongside NeurIPS and ICLR, and its accepted papers tend to show up as shipping features in AI products 12 to 24 months later. That is why it sits on my AI events hub alongside my recap of what actually happened at CVPR 2026 this same season.

This is a recap, not a “should you register” guide - ICML 2026’s main conference wrapped up over a week ago and its workshop track just closed. Here is the real dates and venue, who spoke, who won the paper awards, and what the submission numbers say about where machine learning research is heading. Every claim below traces to ICML’s own conference site and blog or independent technical press, and where I could not verify a number independently, I say so instead of rounding it into something cleaner than it is.

Key Takeaways

• The real dates were July 6-11, 2026, not July 5-10. ICML’s own conference site confirms Monday, July 6 as an Expo and Tutorial day, the main technical program running Tuesday through Thursday, July 7-9, and workshops filling Friday and Saturday, July 10-11. The “July 5-10” date previously attached to this page starts and ends a day earlier than the actual run.
• Submissions hit a record 23,918, more than doubling 2025’s 12,107. Of those, 6,352 papers were accepted, a 26.6% acceptance rate - in line with ICML’s typical low-to-high 20s range even as volume exploded. Within that group, 536 papers (2.2%) earned Spotlight status and just 168 (0.7%) were selected for Oral presentation.
• Two papers shared the Outstanding Paper Award, one challenging assumptions about how diffusion language models generate text, the other closing a theoretical gap in sampling-algorithm efficiency. A 2016 DeepMind paper on asynchronous reinforcement learning - the one that helped make modern RL fine-tuning practical - won the Test of Time Award.
• “Agentic AI” dominated the workshop proposals, not just the buzzwords. ICML’s own workshop chairs flagged that 60-plus submitted workshop proposals used some variant of “agentic AI” in their titles, an unusually high concentration even by ICML’s volume standards, and 44 workshops plus 4 affinity workshops made the final program.
• In-person registration sold out weeks before the conference started, and no official post-event attendance figure exists yet. ICML capped registration in May citing venue and WiFi capacity limits at COEX - a real, dated, sourced constraint - but I could not find a published headcount for the 2026 edition specifically as of this writing.

What Is ICML?

The International Conference on Machine Learning (ICML) is one of the “big three” academic machine learning conferences, alongside NeurIPS and ICLR, organized by the International Machine Learning Society (IMLS) and running annually since 1980. It is a peer-reviewed research venue, not a trade show: the bulk of the program comes from submitted papers that pass double-blind review, supplemented by invited talks, tutorials, workshops, and an expo floor. ICML does not run a single vendor keynote the way NVIDIA GTC does - its “headline moments” are independently scheduled invited talks and paper awards decided by committees the program chairs assemble each year. Compared to NeurIPS and ICLR, ICML has traditionally carried more weight in statistical learning theory, optimization, and reinforcement learning - a positioning that mattered directly in 2026, as this recap gets into below.

ICML’s academic footprint is real and measurable: its proceedings are published open-access through the Proceedings of Machine Learning Research (PMLR), and papers that debuted at past ICMLs - Batch Normalization, EfficientNet, CLIP, Soft Actor-Critic - now sit underneath products most AI teams use daily. If you build tools touching language models or foundation-model training, ICML is where a meaningful share of that research gets published before it shows up in mainstream AI products.

ICML rotates host cities each year. The 2026 edition landed in Seoul; the 2027 edition’s location is due to be announced in August, with 2028 already confirmed for the Eastern United States.

What Happened at ICML 2026: Dates, Venue, and Scale

ICML 2026 ran July 6-11, 2026, at the COEX Convention & Exhibition Center in Seoul, South Korea, per ICML’s own conference site. Monday, July 6 served as the Expo and Tutorial day; the main technical program ran Tuesday through Thursday, July 7-9; and workshops filled Friday and Saturday, July 10-11. That corrects the “July 5-10, 2026” date previously attached to this page, which both starts and ends a day earlier than the conference’s actual run.

Quick facts:

Detail	Information
Event	ICML 2026 (43rd International Conference on Machine Learning)
Dates	July 6-11, 2026 (Expo/Tutorial day Jul 6; main conference Jul 7-9; workshops Jul 10-11)
Venue	COEX Convention & Exhibition Center, Seoul, South Korea
Organizer	International Machine Learning Society (IMLS)
Format	Onsite plus virtual track
Submissions	23,918 (more than double 2025’s 12,107)
Accepted papers	6,352 (26.6% acceptance rate)
Spotlight papers	536 (2.2% of submissions)
Oral papers	168 (0.7% of submissions)
Workshops	44 accepted workshops plus 4 affinity workshops
Invited talks	6 (theory, AI safety, economics, biology, NLP, societal impact)
Outstanding Paper Award	Two winners: diffusion language model reasoning, and log-concave sampling theory
Test of Time Award	“Asynchronous Methods for Deep Reinforcement Learning” (Mnih et al., originally ICML 2016)
Official recap source	ICML’s own dates and venue page (icml.cc), plus the official ICML 2026 awards announcement

The submission growth tells you something about the field, not just the venue. ICML’s own tweet on January 29 confirmed 24,371 submissions crossed the finish line at the full-paper deadline; after desk rejections and author withdrawals, 23,918 remained in the review pool - a figure independently corroborated by both Wikipedia’s citation of ICML’s own tracking and an independent pre-conference technical press writeup covering the same numbers. Either way, that is more than double 2025’s 12,107 submissions, itself a record at the time.

One detail complicates the clean “record year” narrative: 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. Reviewers assigned to the stricter “no LLM use” track had their reviews pulled once detected, and every paper they were responsible for reviewing was desk-rejected regardless of its own quality - an uncomfortable data point about what “record submissions” costs a peer-review system at this scale, and I would rather report it plainly than pretend 2026 was friction-free.

Invited Talks Recap: Six Keynotes Spanning Theory, Safety, and Drug Discovery

ICML 2026’s invited speaker program, announced May 18, brought together six researchers spanning machine learning theory, AI safety and ethics, economics and policy, computational biology, NLP, and human-computer interaction - a deliberately wide net rather than one unifying theme, typical for ICML’s invited-talk slots.

Pascale Fung (Hong Kong University of Science and Technology, co-founder of AMI Labs) spoke on conversational and ethical AI. A Fellow of AAAI, IEEE, ACL, and ISCA, she also sits on the United Nations Advisory Body on AI Governance. Susan Athey (Stanford Graduate School of Business, John Bates Clark Medal recipient) brought the economics of digitization and causal inference into an ML audience. Sham M. Kakade (Harvard’s Kempner Institute) covered reinforcement learning and foundation-model training theory - and is himself a past ICML Test of Time Award recipient, giving his talk a full-circle quality since this year’s Test of Time winner is also an RL paper.

Aviv Regev (Head of Genentech Research and Early Development) addressed AI/ML integration into drug discovery via Genentech’s “Lab in the Loop” approach, reflecting ICML’s growing reach into computational biology. Verena Rieser (Google DeepMind) spoke on responsible development and alignment of frontier AI models. Arvind Narayanan (Princeton, Center for Information Technology Policy) closed the lineup on the societal impact of AI - co-author of AI Snake Oil and a name on TIME’s inaugural 100 Most Influential People in AI list, a fitting closer for a year in which ICML’s own submission numbers forced an uncomfortable conversation about AI’s role in evaluating AI research itself.

Outstanding Papers, the Position Paper Award, and the Test of Time Award

ICML does not use the term “Best Paper” - its top research distinction is the Outstanding Paper Award. Program chairs Alekh Agarwal, Miroslav Dudik, Martin Jaggi, and Sharon Li selected 53 initial candidates from reviewer scores and Area Chair nominations, narrowed that to a 22-paper shortlist, then handed it to an 11-member selection committee chaired by Andreas Krause, which settled on two Outstanding Papers and five Honorable Mentions, per the official ICML 2026 awards announcement.

Outstanding Paper Award, winner one: “The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models.” Zanlin Ni, Gao Huang, and eight co-authors challenge a dominant assumption about diffusion large language models (dLLMs): that generating tokens in arbitrary order is a pure advantage over left-to-right generation. On general reasoning tasks - math and coding - the authors show dLLMs exploit that flexibility to skip exactly the high-uncertainty “forking” tokens that matter most for exploration, collapsing 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.

Outstanding Paper Award, winner two: “High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions.” Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin settle a long-standing open question in score-based sampling theory: 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 over prior results and, as a byproduct, the first polylog(1/ε) gradient-only sampler for general log-concave distributions.

Outstanding Position Paper Award: “Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit.” Sarah Ball and Phil Hackemann argue that alignment methods built to prevent AI harm are dual-use technologies that can be misused for censorship by any authority, present or future - deliberately generalizing the risk rather than naming a country or company. A separate Honorable Mention went to a related piece on AI-generated non-consensual intimate imagery (AIG-NCII) research by Li Qiwei, Wells Lucas Santo, Sarita Schoenebeck, and Eric Gilbert.

Five Outstanding Paper Honorable Mentions rounded out the research track: “The Obfuscation Atlas” (mapping how language models learn to evade deception detectors during RL training), “Motion Attribution for Video Generation” (improved motion quality using one-tenth of the original training data), “How much can language models memorize?” (estimating GPT-style models memorize roughly 3.6 bits of a distribution per parameter), “A Random Matrix Perspective on the Consistency of Diffusion Models” (explaining why models trained on separate data splits produce near-identical outputs from the same seed), and “To Grok Grokking” (the first rigorous bounds on “grokking time” in ridge regression).

Test of Time Award: “Asynchronous Methods for Deep Reinforcement Learning.” Volodymyr Mnih and seven DeepMind co-authors’ 2016 paper - the one that introduced asynchronous actor-critic methods (A3C) - was selected from a shortlist of eight candidates identified by ICML 2016 program co-chair Kilian Weinberger. The committee’s citation credits the paper’s parallel actor-learner insight as “a major contributing factor to the success of RL in LLM post-training” - a direct line from a decade-old paper to how today’s language models get fine-tuned.

Industry Presence and Research Themes at ICML 2026

ICML is not a single-vendor product event, but the workshop proposals told a more concrete story about where the field is heading than the accepted-papers list alone. Per independent technical press coverage published the week the conference opened, ICML 2026 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 accounting for the conference’s volume. 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).

That pivot is not surprising given ICML’s historical center of gravity in optimization theory, statistical learning theory, and reinforcement learning - precisely the toolkit needed to reason about whether an autonomous agent’s policy is safe when it can take irreversible real-world actions. One accepted Oral paper captured the shift concretely: “Do We Need Adam?” found that plain stochastic gradient descent matches or outperforms the widely used AdamW optimizer in the reinforcement learning phase of LLM training, while updating fewer than 0.02% of model parameters - over 1,000 times fewer than AdamW.

On the industry side, Apple sponsored ICML 2026 for at least the sixth consecutive year, staffing a booth across all three main-conference days and presenting more than 20 accepted papers and workshop papers - spanning KV-cache management to diffusion language models to a live demo of “local agentic coding with MLX.” Apple’s own senior area chair list included recognizable industry names, Samy Bengio and Vladlen Koltun among them. Google DeepMind’s presence showed up differently: Verena Rieser gave one of the six invited talks, and DeepMind was the origin institution behind the Test of Time-winning A3C paper from a decade earlier.

Attendance and Scale: How Big Was ICML 2026, Really?

This is a section where I have to be upfront about a gap. ICML’s own blog post from May 24 confirmed that in-person main conference and tutorial registration reached capacity weeks before the conference opened, with organizers citing venue logistics and WiFi infrastructure limits at COEX as the reason for capping registration “marginally below the legal limit for the venue.” ICML Board President Kamalika Chaudhuri framed it this way in the organizers’ own post: “Setting a registration cap is not a decision we take lightly, but it reflects our commitment to making sure that when you come to Seoul, you can actually do the things you came to do.”

That confirms strong demand, but it is not an attendance figure. I could not find an official ICML-published post-event headcount for the 2026 Seoul edition specifically. What I can confirm: 23,918 submissions, 6,352 accepted papers, 536 Spotlight papers, 168 Oral papers, and 44 workshops plus 4 affinity workshops. Submission and acceptance counts are real, 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 plainly that an exact in-person headcount is not yet public than round “sold out registration” into a number I cannot back up.

Was ICML 2026 Worth It? My Honest Take

For the audience ICML is actually built for - machine learning researchers, PhD students, and applied engineers tracking where foundation-model and RL research is heading before it ships - ICML 2026 delivered real substance. Two genuinely different Outstanding Papers, plus a Test of Time Award tracing a straight line from 2016 reinforcement learning research to how 2026’s language models get fine-tuned, is the kind of long-arc validation academic venues are uniquely positioned to provide.

The honest caveat: unless you are actively publishing or reviewing machine learning 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. The desk-rejection controversy over reviewer LLM-policy violations is worth sitting with honestly too - a sign the peer-review system underpinning ICML’s credibility is under real strain at 23,918-submission scale, not a footnote to bury.

If your work touches reinforcement learning, language model training, or foundation-model theory directly, ICML 2026’s paper program and the “agentic AI” workshop concentration are worth tracking closely - the field’s own workshop proposals are telling you where the next two years of research investment are headed. If you build agentic tooling yourself, my breakdown of the best MCP servers tracks the agent-infrastructure ecosystem this kind of research tends to feed into a year or two later, and my roundup of AI literature review tools covers the software side of digging through a 6,352-paper program.

Frequently Asked Questions

When did ICML 2026 actually take place?

ICML 2026 ran July 6-11, 2026, at the COEX Convention & Exhibition Center in Seoul, South Korea - Expo/Tutorial day July 6, main conference July 7-9, workshops July 10-11. This corrects an earlier “July 5-10, 2026” date that did not match ICML’s own official conference site.

What won the Outstanding Paper Award at ICML 2026?

Two papers shared it: “The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models,” led by Zanlin Ni and Gao Huang, and “High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions,” from Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin. ICML does not use the term “Best Paper” - Outstanding Paper Award is its top research distinction.

How many papers were submitted and accepted at ICML 2026?

A record 23,918 submissions reached the final review pool (24,371 crossed the full-paper deadline before desk rejections and withdrawals), more than double 2025’s 12,107. Of those, 6,352 were accepted for a 26.6% acceptance rate, with 536 papers (2.2%) earning Spotlight status and 168 (0.7%) selected for Oral presentation.

How many people attended ICML 2026?

I could not verify an official post-event attendance figure. ICML confirmed in-person main conference and tutorial registration sold out weeks before the conference opened, citing venue and WiFi capacity limits at COEX - a registration-cap disclosure, not a confirmed headcount, and no post-event number has been published for the 2026 Seoul edition specifically as of this writing.

Who gave invited talks at ICML 2026?

Six speakers: Pascale Fung (HKUST, conversational and ethical AI), Susan Athey (Stanford, economics and causal inference), Sham M. Kakade (Harvard’s Kempner Institute, RL and deep learning theory), Aviv Regev (Genentech, AI in drug discovery), Verena Rieser (Google DeepMind, frontier AI alignment), and Arvind Narayanan (Princeton, societal impact of AI).

What happened with the peer-review controversy at ICML 2026?

Program chairs disclosed that 497 papers - about 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. Every paper those reviewers were responsible for reviewing was desk-rejected regardless of its own quality, a direct consequence of enforcing the policy at record submission volume.

The Bottom Line

ICML 2026 ran July 6-11, 2026, in Seoul, not July 5-10 as this page previously stated - and that correction matters less than what happened once the conference opened: a record 23,918 submissions, two genuinely distinct Outstanding Papers spanning empirical and theoretical work, a Test of Time Award tracing a decade-old DeepMind paper’s influence straight into how today’s language models get fine-tuned, and a workshop program where “agentic AI” showed up in more than 60 proposal titles.

Check the primary source - ICML’s own conference site and its awards committee’s writeup - before repeating whatever a stale directory page claims. The submission, acceptance, and award numbers above are real and auditable; the exact in-person attendance figure is not, and I would rather say that plainly than invent a clean-sounding round number.

Next step: if the diffusion-language-model or sampling-theory Outstanding Papers are relevant to what you are building, both are already indexed on arXiv under the “Accepted at ICML 2026” tag. And if you are tracking the broader academic AI events calendar, my CVPR 2026 recap covers the same record-submissions, honest-verdict format for computer vision’s flagship venue. Want a heads-up the moment new AI event coverage goes live? Subscribe here.