ICML 2026 Wraps in Seoul: Record 24,000 Submissions, Agentic AI Everywhere, and a Peer-Review Reckoning

The Forty-Third International Conference on Machine Learning closed out its Seoul edition on Saturday, capping a record-shattering week that confirmed both the field's explosive growth and the strain that growth is putting on its institutions.

The numbers

ICML 2026 drew 23,918 paper submissions — more than double last year — with 6,352 accepted, 536 selected as spotlights and 168 as orals. More than 10,000 researchers converged on the COEX convention center from July 6 to 11, with keynotes from Pascale Fung, Susan Athey and Sham Kakade spanning conversational AI, economics and safety.

Hosting duties gave Korea's AI ecosystem a global stage in a year when the country has poured billions into sovereign AI infrastructure and its chipmakers sit at the center of the AI supply chain.

Agent safety takes over

If one theme dominated the program, it was agentic AI and the safety of autonomous systems. Papers on multi-agent coordination, evaluation-aware behavior, and long-horizon task reliability crowded the orals — a research agenda pulled directly from industry's deployment headaches, from METR's findings on benchmark gaming to enterprise agent rollouts.

Peer review meets its own AI problem

The conference also delivered an uncomfortable milestone: roughly 497 papers — about 2 percent of submissions — were desk-rejected after 398 reciprocal reviewers were found to have violated the conference's LLM usage policies. The episode crystallized a question the field can no longer defer: how a discipline that builds language models should police their use in judging its own work.

ICML's Seoul edition will be remembered as the moment the field's biggest subject — increasingly capable agents — and its biggest institutional problem — reviewing at scale — turned out to be the same story.