← BASELOAD English

ICML 2026 Seoul: AI Rewrote How We Work With It

July 31, 2026 · 42 min read 한국어 원문

ICML — the International Conference on Machine Learning, one of the world’s three premier AI conferences — came to Korea for the first time in 2026, in its official 43rd edition. Held at Seoul’s COEX from July 6 to 11, ICML 2026 drew more than 10,000 researchers and a record-shattering 24,000-plus paper submissions. And the message this vast gathering sent could be summed up in a single line: AI is moving beyond “smarter models” toward “systems that work on their own” — agentic AI.

This piece lays out the overview of ICML 2026, its substance, and what it signals for how we work with AI. The bottom line: the current confirmed in Seoul all points one way — a shift from AI that you instruct step by step, to AI that you hand a goal and it picks its own tools and finishes the job.

Photo by Headway on Unsplash

ICML 2026 Overview — a Milestone: First Time in Korea

ICML ranks alongside NeurIPS and ICLR as one of the world’s top-3 AI conferences — and it’s the oldest of the three. Having begun as a small workshop at Carnegie Mellon University in 1980 and matured into a full conference by 1993, its 43rd edition landing in Asia — specifically Seoul — is itself a testament to Korea’s rising standing in AI research.

This explosive growth is no accident. The 1980 CMU workshop drew only a few dozen people. Since the 2010s deep-learning boom turned machine learning into an industrial engine that soaks up data, compute, and capital, the conference has shifted in character — from a “club for researchers” to an “arena for technological supremacy.” A number like 20,000-plus papers is less a sign of academic maturity than a signal that AI has become the front line of national and corporate strategy.

The scale was unprecedented. The table below shows the headline figures for ICML 2026.

Sources: ICML official (icml.cc); submission statistics from media aggregates (2026).

The submission curve captures the explosion of AI research on its own. From 1,037 papers in 2015 to 12,107 in 2025, the count jumped past 24,000 in 2026 — more than a 20-fold increase in a decade.

Keynotes featured conversational-AI authority Pascale Fung, economist and causal-inference scholar Susan Athey, and AI theorist Sham Kakade, spanning topics from conversational AI to AI safety.

The Dominant Theme — “Agentic AI,” Machines That Work on Their Own

If you had to sum up ICML 2026 in one term, it would be agentic AI. According to the workshop chairs, of the 247 workshop proposals received, 60 contained some variant of “agentic AI” — a concentration described as remarkable even by the conference’s standards.

The research focus has clearly shifted. Where the past few years were a race to build “bigger, smarter models,” the center of gravity in Seoul was “autonomous systems that decide for themselves and work reliably.” Agent training, long-horizon task completion, tool use, and multi-agent coordination filled multiple tracks. The core question became how to build AI that operates dependably — and safely — in the real world without a human in the loop at every step.

In truth, “machines that work on their own” is one of AI’s oldest dreams, and the way it gets realized has changed with each era. In the 1970s–80s, expert systems mimicked “autonomy” by having humans hand-code every rule; in the 1990s–2000s, statistical learning let machines learn the rules from data; in the 2010s, deep learning broke down the wall of perception. Once 2020s LLMs added a layer of “thinking in language,” the remaining piece was agents that “judge and act.” Agentic AI is the newest stratum in this half-century lineage.

And this is exactly where it matters for practitioners. “Agentic AI” means a shift in how we work with AI. We move from asking a chatbot a question, getting an answer, and executing it ourselves — to setting a goal and letting the AI search, call tools, and complete the task on its own. ICML 2026 confirmed that this shift is not marketing spin but a real current at the research frontier.

Photo by Alexandre Pellaes on Unsplash

The second axis was AI safety. The more an agent acts on its own, the more its actions must be predictable and controllable. As autonomy rises, so does the importance of guardrails — which is why the phrase “AI agent safety era” made the rounds. Alongside it, synthetic data generation, to fill gaps in training data, was another major theme.

What the Award Papers Reveal — Diffusion Models and A3C

The year’s research landscape shows up in the award papers, too. Diffusion-model research swept the Outstanding Paper awards. Two were selected: “The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models” and “High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions.” It shows diffusion models — which began in image generation — extending their theoretical reach into language, sampling, and beyond.

Another highlight was the Test of Time Award, given to a paper at least a decade old whose influence has been proven out. It went to DeepMind’s 2016 “Asynchronous Methods for Deep Reinforcement Learning” (the A3C algorithm). By introducing a parallel training technique for reinforcement learning that underpins today’s agent training, it was a fitting, symbolic honor in “the year of agentic AI.”

Korea’s Place — What Naver and LG Showed

As the host country, Korean companies stood out. Team Naver presented research across three axes — from cutting-edge AI models to physical AI: ① strengthening AI safety, ② improving the operational efficiency of models and agents, and ③ 3D spatial understanding and expansion into the physical world. In particular, “Stable-GFlowNet” — selected as a Spotlight — was introduced as a technology that structurally resolves training instability so that a model’s attack vulnerabilities can be verified more robustly and diversely before an LLM is deployed to real services (Digital Today, The Elec).

LG AI Research showcased real-world industrial applications powered by its own large language model, EXAONE. That Korean firms showed up not as spectators but with “full-stack” research (model, safety, application) at the global frontier gave ICML 2026’s Seoul edition meaning beyond the symbolic. Just over a decade ago, ICML and NeurIPS were, to Korean researchers, “someone else’s stage you visit to submit a paper.” Bringing that conference to Seoul and putting a homegrown model (EXAONE) on the front line is a scene that shows a country moving from importing data and infrastructure toward “making the rules.”

So What — a Signal That “How We Work” Is Changing

ICML 2026 leaves us with three implications.

First, the grammar of using AI shifts from “asking” to “delegating.” Agentic AI becoming mainstream at the research frontier means practical tools will soon reorganize in that direction. Going forward, competitiveness will hinge as much on “how to safely delegate work to AI and verify it” as on “how to ask AI good questions.” There’s a déjà vu to this transition. Just as the power loom of the Industrial Revolution, the automation lines of the 20th century, and the scripts of the software era did, the history of tools has always turned “what people operated by hand” into “what runs on its own.” What’s different this time is that what’s being automated is not manual labor or routine tasks, but judgment and decision-making itself.

Second, autonomy and safety are a pair. The more AI acts on its own, the more valuable control and verification mechanisms become. That’s the same logic behind Naver foregrounding “AI safety” and “LLM vulnerability testing.” It’s an era in which, when a company adopts agents, it must ask “how controllable is it?” as much as “how smart is it?”

Third, Korea is moving from consumer to producer. ICML’s first Korean edition and the showings by Naver and LG signal a country shifting from using AI to making it. But for this to be more than a one-off event, it needs sustained backing in talent, investment, and data.

The one line confirmed in Seoul is this: AI has begun to be judged not by “what it knows” but by “what it can do.” And that change will start by reshaping the very way we work with AI, day to day.

Sources

• ICML 2026 official site, icml.cc
• “ICML 2026 Opens in Seoul: Agentic AI Tops Record Year,” TechTimes, techtimes.com
• “Team Naver showcases AI full-stack technology at ICML 2026,” The Elec, thelec.net
• “LG AI Research showcases real-world EXAONE AI applications at ICML 2026,” Korea Times, koreatimes.co.kr