The 43rd International Conference on Machine Learning opened in Seoul, South Korea on Monday, July 6, with a record submission that more than doubles last year’s total and a research agenda defined by questions the field has never had to answer at this scale. It’s about how to manage AI systems that can perform real-world actions without humans directing every step. Registration for the in-person main conference has reached capacity. The 6,352 accepted papers are already available on arXiv for those who want to start reading before the conference begins.
ICML will be held at the COEX Convention & Exhibition Center until Saturday, July 11th, but those who cannot travel to Seoul can still attend virtually.
Record scale, stricter selection
ICML 2026 received 23,918 submissions after initial desk rejections and withdrawals. This is more than double the record number of 12,107 submissions received in 2025. Of these, 6,352 papers were accepted, for an acceptance rate of 26.6%. Within that group, a more selective 536 papers (2.2% of all submissions) earned the Spotlight designation, and only 168 papers, representing the top 0.7% of all submissions, received an oral presentation slot.
This proliferation of posts reflects both the growing size of the global ML research talent pool and the pace at which new research is being produced. At the same time, however, the organizing committee was forced to confront a structural problem: the burden of human review at scale. This year, ICML 2026 program chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, and Martin Jaggi announced that 497 papers, approximately 2% of all submissions, were desk-rejected because they were found to have violated the conference’s LLM usage policy, which 398 peer reviewers had explicitly agreed to abide by. Reviews that violate our policies will be removed. The author lost the paper and any peer review credits they had accumulated.
Agent-based AI becomes a central research topic in machine learning
The clearest signal about where the field was headed did not come from the list of accepted papers. It was born out of a workshop proposal. ICML 2026 workshop chairs Gergely Neu and Courtney Paquette noted in their presentation that several variations of “Agent AI” appeared in the titles of the more than 60 workshop proposals submitted, a concentration they said was surprising considering the conference volume.
Agenttic AI refers to a system that receives a goal, decides what steps to take, calls a tool or API, observes the results, and continues without the need for human approval at each step in between, pursuing the goal through its own autonomous loop of action. The reason this research direction has received so much attention, especially in ICML (a venue that has historically focused on optimization theory, statistical learning theory, and reinforcement learning), is that the issue of safety and reliability of autonomous agents lies precisely at the intersection of these fields. When agents can perform irreversible actions in the real world, the question of whether a policy is safe is no longer an academic one. The final workshop program accepted 44 workshops and 4 affinity workshops to be held concurrently with the main conference.
A workshop was approved to directly address this change. Among them is the second edition of Agents in the Wild, which focuses on safety, security, and multi-agent coordination in open-end environments. “A statistical framework for uncertainty in agent systems.” It brings together researchers working on conformal prediction and calibration of agent pipelines. “AI for Science: AI Scientists—Tools, Coauthors, or Founders?” considers what it means for autonomous systems to conduct rather than support research.
What oral papers reveal about technical priorities
The 168 oral presentations provide the clearest understanding of what the ICML Program Committee has determined to be the most important contributions to the field. One stands out for challenging deep assumptions about how large-scale language models are trained. The paper “Do we need Adam?” found that SGD (a classic stochastic gradient descent algorithm that predates almost all modern neural network training) performs as well as or better than the widely used AdamW optimizer, especially in the reinforcement learning phase of LLM training, updating less than 0.02% of model parameters, more than 1,000 times less than AdamW. The engineering implications are important. RL fine-tuning of large models can potentially be much more memory efficient than assumed in current practice.
The Oral list also includes work across stochastic density and score estimation using genome-based models, minimax optimization in a game-theoretic setting, and the Transformer architecture, which recovers normalized kernel density estimation as a special case of self-attention. This technology scope reflects ICML’s historical position relative to peer conferences. In other words, compared to NeurIPS and ICLR, ICML has consistently placed more emphasis on optimization theory, statistical learning theory, and reinforcement learning. These are the same areas that are currently most urgently concerned with increasing the reliability of autonomous systems.
Six keynote speeches spanning AI safety, drug discovery, and economic theory
The lineup of invited speakers spans six research communities. pascal fan The Hong Kong University of Science and Technology professor, who is a fellow of AAAI, IEEE, ACL and a member of the United Nations Advisory Council on AI Governance, will speak on conversational AI and ethical AI. susan asayProfessor of Technology Economics at Stanford University Graduate School of Business and recipient of the John Bates Clark Medal, he brings a perspective on causal inference at the intersection of AI and economic systems. Shyam M. Kakade A recipient of the ICML Test of Time Award, the PhD from the Kempner Institute at Harvard University covers theoretical and practical advances in deep learning, reinforcement learning, and basic model training.
Aviv RegevGenentech, head of research and early development, will address the integration of AI and ML in drug discovery. This combination reflects the expansion of ICML into computational biology. Verena Reaser Google DeepMind talks about tuning and responsible development of frontier AI. A sixth speaker on human-computer interaction rounds out the lineup.
Program structure and virtual access
Conference week will be held as follows: Monday, July 6th is Expo and Tutorial Day, featuring hands-on sessions, panel discussions, and industry demonstrations. The main conference will be held Tuesday through Thursday from July 7th to 9th and will include oral presentations, spotlight poster sessions, position paper posters, and journal track posters. The last two days of the workshop will be held on Friday and Saturday, July 10th and 11th.
Registration for the workshop is still being accepted. Authorized contributors can register even if they haven’t registered yet. Virtual registration, required for authors who do not present in-person, remains available and is not subject to in-person capacity limits.
Complete conference details, schedule planner via Scholar Inbox, and complete list of accepted papers are available at icml.cc. The paper will be published through PMLR (Proceedings of Machine Learning Research).
FAQ
How can I access ICML 2026 papers without attending the conference in person?
All 6,352 accepted papers will be available to read before the conference begins. Search arXiv using “Accepted at ICML 2026” as a filter tag, or see the complete list of papers at icml.cc/virtual/2026/papers.html. Virtual registration, which is not subject to in-person capacity limitations, is required for authors presenting remotely, but all registrants will be granted access to the virtual program. The minutes of the meeting will be made available for open access through PMLR after the meeting.
What does the dominance of agent AI in the ICML Workshop proposal mean for machine learning research?
The fact that more than 60 workshop proposals use “Agent AI” in their titles suggests a structural shift, not just a trend. The field is shifting its primary focus from building more capable models to building autonomous systems that can operate reliably and safely in the real world without step-by-step human supervision. This is important for ICML, especially since the technical issues of agent safety (optimizing constrained policies, quantifying uncertainty in multi-step decision systems, conformal guarantees for agent pipelines) are directly integrated into the optimization theory and reinforcement learning research tradition that ICML has historically focused on. The questions that ICML 2026 is posing are different in kind than those posed by ICML five years ago.
When will the ICML 2026 main conference be held and what are the workshops?
The main conference will be held from Tuesday, July 7th to Thursday, July 9th at the COEX Convention & Exhibition Center in Seoul. Monday, July 6th is Expo and Tutorial Day. The workshop program will take place on Friday and Saturday, July 10th and 11th, with 44 accepted workshops and four affinity workshops running concurrently with the main conference dates. Registration for the in-person main conference has reached capacity, but registration for the workshops is still being accepted as of today.
Why did ICML 2026 desk-reject nearly 500 papers before the review process was complete?
The desk rejection occurred during the review, not before. ICML 2026 implemented two track policies governing whether reviewers can use LLM to evaluate papers. After the review period began, the program chair used watermarking technology to detect reviewers who violated the policy track they had agreed to follow. 398 reviewers who violated Policy A (prohibiting the use of LLM) had their reviews removed. All 497 papers for which I was a peer reviewer were rejected on paper, regardless of their content or quality. This incident exposed structural tensions in large-scale peer review. As the volume of submissions doubles, the workload of reviewers is increasing faster than the human reviewer pool can absorb it.
