ICML to host workflow and safety workshops

The International Conference on Machine Learning will host 44 workshops in Seoul in July 2026, including sessions on AI research workflows and agent safety. (icml.cc ) ICML 2026 runs July 6 to July 11 at the Coex Convention and Exhibition Center in Seoul, South Korea. The main conference is July 7 to July 9, and workshops run July 10 and July 11. (icml.cc ) One of the new sessions is “AI as a Tool for Mathematics, Computer Science, and Machine Learning,” listed on OpenReview as an ICML 2026 workshop scheduled for July 10, 2026. Its submission deadline is May 14, 2026, at 11:59 a.m. Coordinated Universal Time. (openreview.net ) A second session, “Agents in the Wild: Safety, Security, and Beyond,” is planned for July 10 or July 11, 2026. Its organizers say the workshop will cover agent safety, security, privacy, robustness, hallucination, factuality, interpretability, fairness, benchmarking, and long-horizon behavior. (agentwild-workshop.github.io ) In plain terms, a workshop is the part of a conference where researchers test early ideas, compare methods, and argue about open problems before a field settles on standard practice. ICML’s own call says workshops are meant to discuss emerging questions and challenges, with one-day formats that leave room for posters, panels, and debate. (icml.cc ) That format matters in 2026 because ICML says workshop selection was unusually competitive this year. The conference received 247 workshop proposals, up from about 150 for ICML 2025, and accepted 44. (blog.icml.cc ) The research-workflow side is already branching beyond one session. A separate “AI for Math” workshop, now in its third year, is billed as “Toward Self-Evolving Scientific Agents” and asks how systems can support theorem proving, autoformalization, scientific problem solving, and measurement while staying verifiable and reliable. (ai4math2026.github.io ) The workshop page describes formal theorem proving as getting a computer to check each logical step of a proof, like a line-by-line calculator for mathematics. It says recent language-model and formal-methods advances have pushed systems to stronger theorem-proving and natural-language math performance, but frames verification and reliability as central constraints. (ai4math2026.github.io ) ICML’s own conference page shows the same pressure spilling into conference policy. The site says ICML 2026 has announced a policy for large language model use in reviewing, alongside posts about an experimental paper-assistant program and violations of review rules. (icml.cc ) By July, the question in Seoul will be less whether researchers use AI tools than how they measure, constrain, and compare them. ICML’s workshop slate suggests those arguments are moving from side conversations into the conference program itself. (blog.icml.cc )