ICML 2026 Kicks Off in Seoul: Record 23,918 Submissions, Agentic AI Dominates as Peer Review Crisis Unfolds

SEOUL, South Korea — The preeminent International Conference on Machine Learning (ICML 2026) commenced today at the COEX Convention & Exhibition Center in Gangnam, Seoul, marking a watershed moment in the 43-year history of the world's largest machine learning conference icml.cc . More than 11,000 researchers have converged for a six-day event running through July 11, 2026, navigating unprecedented challenges that extend far beyond the technical presentations www.techtimes.com .

The deluge of submissions reflects machine learning's explosive growth as a discipline, but it has also exposed critical lacunae in the peer review infrastructure that governs scientific validity. For the first time, ICML arrives carrying a peer-review integrity crisis as significant as any paper in its program—a structural breakdown of the human review process occurring simultaneously with the technical acceleration of autonomous AI systems.

Prompt Injection Exposes 398 Reviewers Violating LLM Ban

The most technically significant institutional event to emerge from ICML's preparation involves an ingenious enforcement mechanism that caught 398 reviewers using large language models they had explicitly agreed not to use. The program committee employed a technique based on recent research by Rao, Kumar, Lakkaraju, and Shah, creating a dictionary of 170,000 phrases and sampling two phrases randomly for each submitted paper.

The family-wise error rate—the probability of incorrectly flagging even a single Policy A review—was calculated at 0.0001, making this one of the most rigorous academic integrity enforcement operations in conference history. Every flagged review underwent manual inspection by a member of the organizing committee before sanctions were imposed.

Agentic AI's Structural Arrival Dominates Workshop Proposals

Beyond the integrity crisis, the clearest signal about where machine learning research is heading emerged from 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—a volume they described as noteworthy even by ICML's standards www.techtimes.com . 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 human approval for each intermediate step.

Among the accepted workshops directly addressing this paradigm shift: 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.

Opening Day Highlights: Seoul World Model and Multi-Agent Systems

Monday, July 6—designated as Expo and Tutorial Day—featured a panoply of presentations showcasing the frontier of machine learning research icml.cc . The morning session opened with an Expo Talk Panel on "Multi-Agent System Design and Evaluation for Quantitative Finance" by Lucas Baker and Loren Puchalla Fiore from Jump Trading, presenting results from firmwide benchmarks evaluating multi-agent architectures operating under extreme constraints across thousands of instruments and terabytes of daily market data.

The afternoon featured multiple demonstrations, including Google's presentation on "Pushing the Frontiers of Large-Scale 3D Modeling for Robotics & Beyond" by Krzysztof Choromanski, and the unveiling of Fara1.5, a family of lightweight Computer Use Agent models at three scales (4B, 9B, and 27B) achieving state-of-the-art results on browser-use benchmarks among models of comparable size.

Oral Papers Challenge Foundational Assumptions

The 168 Oral presentations represent what the ICML program committee judged to be the field's most significant contributions. One paper stands out for challenging a foundational assumption in how large language models are trained: "Do We Need Adam?" finds that stochastic gradient descent (SGD)—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.

Key Finding: The research demonstrates that SGD updates fewer than 0.02% of model parameters while maintaining or improving performance—more than 1,000 times fewer than AdamW—suggesting that 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. 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 whether AI systems can be steered toward specific value frameworks rather than a single fixed alignment target.

Community Building Through Social Events

Social Chairs Kevin Leyton-Brown and Chulhee Yun have curated 12 socials running throughout the conference, covering topics from open-source research to chess to networking at conferences blog.icml.cc . Monday's socials include "AI Co-scientists in the Research Loop," inviting attendees to share practices and discuss emerging norms around disclosure, authorship, reproducibility, and the responsible use of AI agents in research—a direct response to the peer review crisis that has dominated pre-conference discourse.