ICML 2026 Accepts Over 6,500 Papers as Seoul Conference Showcases Largest Machine Learning Research Program Ever

--- headline: "ICML 2026 Accepts Over 6,500 Papers as Seoul Conference Showcases Largest Machine Learning Research Program Ever" slug: icml-2026-6500-papers-seoul category: research story_number: "11" date: 2026-05-08 sources: - name: ICML 2026 Official Conference url: https://icml.cc/Conferences/2026 domain: icml.cc - name: Paper Digest - ICML 2026 Highlights url: https://resources.paperdigest.org/2026/05/icml-2026-papers-highlights/ domain: resources.paperdigest.org - name: KAIST SAIL Lab url: https://sail.kaist.ac.kr/paper-a-paper-on-reliable-feature-attribution-is-accepted-at-icml-2026-2/ domain: sail.kaist.ac.kr - name: Aarhus University Department of Computer Science url: https://cs.au.dk/news-events/news/show-news/artikel/major-success-at-icml-2026-with-13-accepted-papers-spotlight-recognition-and-tutorial domain: cs.au.dk - name: ICML Blog - LLM Review Policy Violations url: https://blog.icml.cc/2026/03/18/on-violations-of-llm-review-policies/ domain: blog.icml.cc ---

The International Conference on Machine Learning has shattered its own records. ICML 2026, scheduled for July 6-11 at the COEX Convention & Exhibition Center in Seoul, South Korea, accepted 6,352 papers out of 23,918 submissions that entered the review process — a 26.6% acceptance rate that will make its 43rd edition the largest program in the conference's four-decade history.

The raw numbers alone tell a story of a discipline in hypergrowth. Total submissions reached 24,371 before desk rejections and withdrawals, more than doubling the 12,107 papers submitted to ICML 2025. That near-100% year-over-year increase represents the steepest single-year jump in the conference's history and reflects the continued explosion of machine learning research worldwide, fueled by both academic expansion and industrial AI investment.

A Global Research Showcase in Asia's AI Capital

The decision to host ICML 2026 in Seoul carries symbolic weight. South Korea has rapidly emerged as a global AI powerhouse, home to institutions like KAIST, Seoul National University, and corporate research labs at Samsung and LG that are contributing to the field at an accelerating pace. General Chair Tong Zhang of the University of Illinois leads the organizing effort alongside Program Chairs Miroslav Dudik of Microsoft Research, Martin Jaggi of EPFL, Alekh Agarwal of Google, and Sharon Li of the University of Wisconsin-Madison.

The six-day program follows a structured format: an Expo and Tutorial Day on July 6, the main conference running July 7-9, and workshops on July 10-11. All accepted papers were reviewed through a double-blind process, and authors can choose to present in person or opt for proceedings-only inclusion — though all accepted papers receive equivalent treatment in the proceedings and remain eligible for ICML awards and distinction designations.

Among the research highlights, KAIST's Statistical Artificial Intelligence Lab secured acceptance for a paper on manifold-aligned guided integrated gradients for reliable feature attribution, addressing how AI systems explain their own decisions — a topic of growing importance as machine learning models are deployed in high-stakes domains. Aarhus University's Algorithms, Data, and Artificial Intelligence section celebrated 13 accepted papers, including a Spotlight designation for a paper on revenue efficiency of correlated equilibria in first-price auctions by Anders Bo Ipsen and Stratis Skoulakis.

"Researchers from our section have achieved a major research milestone at the 43rd International Conference on Machine Learning," Aarhus University stated in its announcement, noting that the Spotlight designation is "awarded to a small selection of particularly notable contributions."

The LLM Elephant in the Review Room

Perhaps the most consequential development at ICML 2026 had nothing to do with the research itself but rather with how it was evaluated. The conference implemented what observers have called the strictest AI integrity policies in the history of academic machine learning. LLMs were explicitly banned from being listed as authors, and any attempt at prompt injection — embedding text designed to manipulate AI-assisted review tools — resulted in automatic desk rejection.

ICML 2026 introduced a novel two-policy framework for reviewer use of AI tools. Under Policy A, the conservative option, reviewers agreed to forgo LLM assistance entirely. Under the more permissive Policy B, reviewers could use LLMs to help understand papers and polish their reviews, but were prohibited from asking AI systems to identify strengths and weaknesses or draft review content.

The enforcement mechanism was as inventive as it was aggressive. Conference organizers watermarked submission PDFs with hidden instructions that would trigger specific phrases if a reviewer fed the paper to an LLM. The detection system identified 795 reviews from 506 unique reviewers who had pledged to follow Policy A but violated their commitment.

"The probability of incorrectly flagging even a single submitted review was 0.0001," the ICML organizing committee stated in a blog post detailing the enforcement results. The consequences were severe: 497 papers were desk-rejected as a penalty for their authors' violations of review integrity agreements, corresponding to submissions from 398 reciprocal reviewers.

What the Numbers Signal

The doubling of submissions raises questions that the acceptance rate alone cannot answer. While a 26.6% acceptance rate is only marginally lower than 2025's 26.9%, the absolute volume of accepted work — 6,352 papers — creates practical challenges for attendees trying to navigate the program and for the research community trying to absorb the output.

The conference also featured new additions to address this scale problem. Aarhus University's Akhil Arora and Nouha Dziri were selected to present a tutorial on adaptive reasoning in large language models, covering the pipeline from post-training to test-time learning — a topic that sits at the center of current industrial AI development.

As machine learning conferences grow to the scale of small cities, ICML 2026 in Seoul represents both a celebration of the field's vitality and a stress test of its institutions. The 6,352 accepted papers will generate thousands of conversations in the halls of COEX this July. Whether the community's review infrastructure can keep pace with its ambition may be the most important research question of all.