The world’s largest machine learning conference opens Monday in Seoul, South Korea — and for the first time in the 43rd-year history of the International Conference on Machine Learning, it arrives carrying a peer-review integrity crisis as significant as any paper in its program. The 11,000-plus researchers descending on the COEX Convention & Exhibition Center through July 11 will navigate a week shaped by two forces that the field has never had to manage simultaneously at this scale: the technical acceleration of autonomous AI systems, and the structural breakdown of the human review process that decides which research gets to be called science.
On the same day ICML opens, the inaugural UN Global Dialogue on AI Governance convenes at the Palexpo International Convention Centre in Geneva — the first intergovernmental forum mandated by the UN General Assembly to address international AI cooperation, running July 6–7. The juxtaposition is not coincidental. The researchers filling COEX’s poster halls this week are producing the work that policymakers in Geneva are trying to govern. The gap between those two processes — how fast the research is moving, and how slowly governance assembles — defines what is actually at stake in both rooms.
Record Volume, Tighter Selection
ICML 2026 received 23,918 submissions after an initial round of desk rejections and withdrawals — more than double the 12,107 submissions the conference received in 2025, which had itself set a prior record. Of those, program chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, and Martin Jaggi accepted 6,352 papers at a 26.6% acceptance rate. Within that group, 536 papers earned Spotlight designations (2.2% of all submissions) and just 168 received Oral presentation slots — the top 0.7% of everything submitted.
All 6,352 accepted papers are pre-published and available to read before the conference opens. Researchers can browse the full list at icml.cc/virtual/2026/papers.html or filter arXiv using the tag “Accepted at ICML 2026.” Proceedings will be published open-access through PMLR (Proceedings of Machine Learning Research) following the conference. Virtual registration — not subject to the in-person capacity cap — remains available for those who cannot travel to Seoul.
How Prompt Injection Caught 398 Reviewers Using LLMs They Had Agreed Not to Use
The peer-review enforcement story is the most technically significant institutional event to emerge from ICML’s preparation for 2026, and it is worth explaining precisely how it worked — because the method reveals as much about the limits of enforcement as it does about the scale of the problem.
ICML implemented a two-track LLM policy for reviewers. Policy A prohibited all LLM use in review writing. Policy B permitted limited LLM assistance — understanding a paper, polishing prose — but not judging quality or drafting the review. Reviewers chose their track. The only reviewers assigned to Policy A were those who explicitly selected it, or who selected “either policy.” No reviewer who strongly preferred Policy B was assigned to Policy A.
To detect violations, the program committee used a technique based on recent research by Rao, Kumar, Lakkaraju, and Shah (published in PLOS ONE in September 2025 as “Detecting LLM-generated peer reviews”). Program chairs created a dictionary of 170,000 phrases and sampled two phrases randomly for each submitted paper. The probability of any given pair being selected was smaller than one in ten billion — making coincidental reproduction essentially impossible. Each submitted PDF was then modified to include machine-readable instructions, invisible to a human reader but visible to any LLM processing the document, instructing the LLM to include both phrases in the review it generated.
In pre-deadline tests, frontier LLMs followed the injected instructions more than 80% of the time. Every flagged review was manually inspected by a member of the organizing committee. The family-wise error rate — the probability of incorrectly flagging even a single Policy A review — was calculated at 0.0001.
The result: 795 reviews (approximately 1% of all reviews submitted) were flagged as LLM-generated by Policy A reviewers who had explicitly agreed not to use LLMs. Of those, 506 unique reviewers were caught. The 398 who were “reciprocal reviewers” — meaning their submission to ICML depended on their serving as a reviewer — had their papers desk-rejected: 497 papers in total, roughly 2% of all submissions.
But the program chairs acknowledged the method’s limits plainly. Inserting hidden instructions in submitted PDFs “is not a difficult measure to circumvent,” they wrote on the ICML blog, “particularly if it is known publicly — which was the case for almost the entire review period.” The method, by design, can only catch reviewers who feed the entire PDF to an LLM and copy-paste the output directly. Sophisticated users who paraphrase, use LLMs for partial drafts, or work around the injected prompt would not be detected.
That constraint carries an implication the raw numbers do not state: if 1% of reviews were caught using a detection method that by the organizers’ own admission catches only the most careless users, the actual rate of LLM use in Policy A reviews is likely substantially higher. “I have been working on conference peer review for several years, and I have hardly seen such strong support for anything,” scientific integrity chair Nihar Shah of Carnegie Mellon University told The Transmitter. “People were really tired of reviewers copy-pasting AI-generated reviews without putting any effort.”
Not everyone welcomed the approach. Researcher Sören Auer told The Transmitter he considered hidden prompts a problematic enforcement mechanism, adding that “it’s not good to prohibit the use of AI — we should rather have a discussion on how to use it.” Sara Atito of the University of Surrey called the technique a “poor mechanism” that filters some violations without addressing the structural problems in peer review. NeurIPS, which holds its annual conference in Sydney in December 2026, has independently adopted similar prompt-injection enforcement and declined to discuss details publicly for fear of undermining the method’s effectiveness.
Agentic AI’s Structural Arrival in Machine Learning Research
The clearest signal about where ML research is heading did not come from the accepted papers list. It came from the workshop proposals. Workshop chairs Gergely Neu and Courtney Paquette noted that some variation of “agentic AI” appeared in no fewer than 60 of the 247 workshop proposals received — a volume they described as remarkable even by ICML’s standards. The final program accepted 44 workshops from those 247 proposals, plus 4 affinity workshops running concurrently with the main conference days.
Agentic AI refers to systems that pursue goals through autonomous action loops: receiving an objective, selecting a course of action, calling tools or APIs, observing the results, and iterating — without requiring a human to approve each intermediate step. OpenAI’s Operator product and open-source tools like OpenClaw represent early commercial deployments of this architecture.
The reason agentic AI is dominating ICML specifically — a venue historically weighted toward optimization theory, statistical learning theory, and reinforcement learning — is that the safety and reliability problems of autonomous agents sit precisely at the intersection of those fields. When an agent can take irreversible actions in the real world, the question of whether its policy is safe is no longer a theoretical concern. The technical research problems that follow — constrained policy optimization, uncertainty quantification for multi-step decision systems, conformal prediction guarantees for agentic pipelines — are exactly the problems ICML has the deepest research tradition to address.
Among the accepted workshops directly addressing this: a second iteration of “Agents in the Wild,” focused on safety, security, and multi-agent coordination in open-ended environments; “Statistical Frameworks for Uncertainty in Agentic Systems,” addressing conformal and calibration methods for agent pipelines; and “Technical AI Governance Research,” a policy-meets-ML bridge workshop covering formal approaches to governing AI development.
What 168 Oral Papers Reveal About the Field’s Technical Priorities
The 168 Oral presentations represent what the ICML program committee judged to be the field’s most significant contributions. One stands out for challenging a foundational assumption in how large language models are trained.
The paper “Do We Need Adam?” finds that SGD — stochastic gradient descent, the classical optimization algorithm that predates nearly all modern neural network training methods — matches or outperforms the widely used AdamW optimizer specifically during the reinforcement learning fine-tuning phase of LLM training, while updating fewer than 0.02% of model parameters — more than 1,000 times fewer than AdamW. The engineering implication is material: RL fine-tuning of large models may be substantially more memory-efficient than current practice assumes.
The Oral list also spans work on genomic foundation models, game-theoretic minimax optimization, and transformer-based probabilistic density estimation — a cluster of papers that recover normalized kernel density estimation as a special case of self-attention. The alignment and safety track contains several closely watched papers, including “The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes” and “VALUEFLOW: Toward Pluralistic and Steerable Value-based Alignment in Large Language Models” — research directly addressing the question of whether AI systems can be steered toward specific value frameworks rather than a single fixed alignment target.
Six Keynotes Span AI Safety, Drug Discovery, and Economic Theory
The invited speaker lineup intentionally crosses research communities in a way recent ICML lineups have not. Six researchers were selected — announced in May by communications chair Gautam Kamath — whose work spans well beyond traditional ML systems:
Pascale Fung of the Hong Kong University of Science and Technology opens the main conference on Tuesday. A Fellow of AAAI, IEEE, ACL, and ISCA, and a member of the United Nations Advisory Body on AI Governance, Fung speaks on “Towards AI Agents in the Real World” — building on her 2025 ACL alignment keynote with a human-centered roadmap for the transition from instruction-following systems to principled autonomous agents.
Susan Athey, Economics of Technology Professor at Stanford’s Graduate School of Business and the first woman to receive the John Bates Clark Medal, brings a formal causal inference perspective to AI’s distributional effects — what the technology does to economic institutions, labor markets, and the people who work within them.
Sham M. Kakade, Rampell Family Professor at Harvard and co-director of the Kempner Institute, covers the statistical and mathematical foundations of reinforcement learning and deep learning, including theory for the feature learning and memorization phenomena that shape how foundation models generalize.
Aviv Regev, Executive Vice President and Head of Genentech Research and Early Development, presents her lab’s work integrating ML into drug discovery through a “Lab in the Loop” approach — closed-loop experimental design where ML systems guide biological experiments rather than just analyze their results.
Verena Rieser of Google DeepMind, a Leverhulme Trust Senior Research Fellow and honorary professor at Heriot-Watt University, presents a responsible development roadmap for frontier AI alignment — combining conversational reinforcement learning agents with sociotechnical frameworks for principled behavior.
Arvind Narayanan, Professor of Computer Science at Princeton and director of the Center for Information Technology Policy, speaks on the science of AI agent evaluation and AI’s impact on institutions and professions. Named to TIME’s inaugural list of 100 most influential people in AI, Narayanan co-authored AI Snake Oil and the widely-read essay “AI as Normal Technology.”
The inclusion of an economist (Athey), a computational biologist (Regev), and a computer scientist known primarily for policy critique (Narayanan) marks a deliberate broadening of ICML’s historically theory-heavy invited speaker tradition — a signal that the conference is treating AI governance and distributional effects as problems that belong on the research agenda, not just in side conversations.
Schedule and Access
ICML 2026 runs July 6–11 at the COEX Convention & Exhibition Center in Gangnam, Seoul. Monday, July 6 is Expo and Tutorial Day, with ten 2.5-hour tutorial sessions covering topics from machine unlearning at production scale to Kan Extension Transformers and neural theorem proving with the Lean proof assistant. The main conference runs Tuesday through Thursday, July 7–9. Workshop days fill Friday and Saturday, July 10–11.
Registration for tutorials and the main conference in-person track is sold out. Workshop registration remains open. Virtual registration, available for any registrant, is not subject to the in-person cap and provides access to the virtual program and proceedings. All accepted papers are currently available at icml.cc/virtual/2026/papers.html and on arXiv.
Tickets range from $50 for workshop-only access to $1,365 for full industry registration. Registration opened Sunday, July 5, with a desk at COEX from 2:00 PM KST. Seoul’s Samseong station on Line 2 provides direct subway access to the venue. July weather in Seoul is humid and prone to afternoon rain — the final stretch of the East Asian monsoon season.
In Seoul and Geneva, Two Conversations About the Same Problem
Pascale Fung, who opens ICML 2026 with Monday’s keynote, holds a position that few researchers in the world occupy: she serves simultaneously on the United Nations Advisory Body on AI Governance and leads one of Asia’s leading conversational AI research groups. Her presence as the conference’s opening speaker is a deliberate editorial choice by the program committee — a statement that the field’s technical and governance conversations have converged to the point where they cannot be held in separate rooms.
In Geneva on July 6, delegations from 193 UN member states are convening to begin building the first multilateral governance framework for AI — a process that will eventually address access, safety, accountability, and AI sovereignty for every nation that lacks the capacity to build foundation models itself. In Seoul, researchers are presenting the peer-reviewed technical work that makes those governance conversations both urgent and structurally incomplete: new agent architectures, alignment failure modes, interpretability methods, benchmark exploits, and the continued scaling of systems that neither diplomats nor engineers fully understand.
The distance between the technical frontier and the governance framework is not a failure of either institution. It is the defining structural condition of AI development in 2026 — and understanding that gap is what makes both events worth following this week.
Frequently Asked Questions
How did ICML 2026 detect reviewers who used LLMs when they weren’t supposed to?
Program chairs modified each submitted PDF to include machine-readable instructions — invisible to human readers but interpretable by LLMs — telling any AI processing the document to include two specific phrases in the review. The phrases were drawn randomly from a dictionary of 170,000 entries; the probability of any pair appearing coincidentally is smaller than one in ten billion. Every flagged review was manually verified by a human. Frontier LLMs followed the injected instructions more than 80% of the time in pre-deadline tests. The method only catches reviewers who feed the full PDF directly to an LLM and copy-paste the output; the organizers acknowledged it would not detect more sophisticated users. The full technical explanation is in the ICML Peer Review FAQ at icml.cc.
What does agentic AI’s dominance in ICML 2026 workshop proposals mean for researchers outside ML?
It signals that autonomous AI systems — ones that pursue goals through their own action loops without step-by-step human direction — are no longer a research prototype but an active engineering and deployment problem. For researchers in adjacent fields like economics, biology, law, and policy, the practical implication is that the systems these researchers will interact with professionally are changing from tools that respond to prompts to agents that make decisions and take actions. ICML is where the safety guarantees, uncertainty bounds, and evaluation frameworks for those systems are being developed — or found to be missing.
If the watermarking method is easily bypassed, does it actually solve the peer review problem?
No — and the ICML organizers said as much. The method is acknowledged to catch only the most careless users: those who feed a PDF directly to an LLM and copy-paste the output without modification. Sophisticated reviewers who paraphrase, use LLMs for partial drafts, or work around the injected prompt would not be flagged. The 1% detection rate should therefore be read as a floor estimate of LLM misuse, not a ceiling. What the enforcement action did accomplish is establishing a precedent: that conferences can take structural action against violations, that the scale of violation is measurable, and that the “honor system” model of peer review is functionally insufficient at current submission volumes. NeurIPS has adopted a similar approach for its December 2026 conference.
Can I still attend ICML 2026 workshops if I didn’t register for the main conference?
Yes. In-person main conference registration is sold out, but workshop registration remains open and is not subject to the in-person capacity cap. Virtual registration — which provides access to the virtual program and all session recordings — is also still available. Authors of accepted posters who have not yet registered can still do so. Workshop registrations and virtual registration are processed at icml.cc/Register2.
