ICML 2026 Opens in Seoul: Agentic AI Dominates a Record-Breaking Year

TL;DR: ICML 2026 opened today at COEX Seoul with 23,918 submissions — more than double 2025’s record. Agentic AI appears in 60 of 247 workshop proposals, 749 of 6,796 papers are agent-relevant, and the conference’s keynote lineup signals a field-wide pivot toward safety, reliability, and real-world deployment. Here’s what the numbers say and which papers you need to know.

The Scale of ICML 2026

The 43rd International Conference on Machine Learning opened its doors today at the COEX Convention & Exhibition Center in Seoul’s Gangnam district — the first time ICML has landed in South Korea. The headline number stopped the field in its tracks: 23,918 submissions, more than double the 12,107 that set a record just one year earlier. From that pool, 6,352 papers were accepted (26.6% acceptance rate), representing roughly 25,000 authors.

The two-year trend is stark. For comparison, ICML 2024 received ~9,600 submissions. The field has seen 2.5× growth in two years — and the composition of those submissions has shifted dramatically.

(Source: ICML 2026 — Official Program)

Agentic AI Takes Center Stage

When workshop chairs tallied the 247 proposals they received, some variation of “agentic AI” appeared in no fewer than 60 of them. The Cortiq agentic AI map for ICML 2026 counts 749 agent-relevant papers out of 6,796 total — roughly 11% of the program — alongside 12 agentic-themed workshops and 51 exhibiting companies.

For a venue historically weighted toward optimization theory and statistical learning, that concentration is a structural shift. The field’s center of gravity is moving from models that predict to systems that act.

(Source: Cortiq — ICML 2026 Agentic AI Map)

Keynote: Pascale Fung on “Towards AI Agents in the Real World.” Fung, who serves on the United Nations Advisory Body on AI Governance and co-founded AMI Labs, opened the conference positioning agent safety not as an ethics sidebar but as a core research problem. Her framing: agents trained via imitation learning and RL in digital environments deliver strong online task performance, but their capability in physical-world settings remains limited — and the gap is where governance becomes urgent.

(Source: AI2Work — ICML 2026 Opens in Seoul)

The Safety Imperative: FAGEN and the Failure Mode Research Agenda

The most telling workshop for anyone deploying agents in production: FAGEN (Failure Modes in Agentic AI), running July 10. Its stated deliverables are a research program in themselves:

Operational definitions with explicit boundaries and loop localization
Minimal, reproducible triggers for agent failures
Comparable protocols with trace-level diagnostics
Verified fixes

Three contributed papers illustrate the range: D-CEM challenged unsafe consensus in multi-agent deliberation, Who&When Pro benchmarked multimodal failure attribution at scale, and ATLAS showed how adaptive failure taxonomies can improve judging, reflection, and agent optimization.

This is a notable maturation. A year ago, the conversation was whether agents could work at all. Now ICML has an entire workshop dedicated to classifying and systematically fixing their failure modes.

(Source: FAGEN Workshop @ ICML 2026)

SCALE (Scalable Learning and Optimization for Efficient Multimodal AI Agents), also running July 10, tackles the complementary problem: how to make multimodal agents efficient enough to operate at production scale without sacrificing reliability. The agenda spans pretraining pipelines, fine-tuning strategies, and test-time adaptation — all through the lens of deployment viability.

(Source: SCALE Workshop @ ICML 2026)

Five Papers That Define Where Agent Research Is Going

One research team at ICML 2026 has five accepted papers that together map the current agent research stack. Here they are:

InfoPO: When Should an Agent Ask Questions?

User requests are often underspecified — missing constraints, implicit preferences, ambiguous goals. InfoPO introduces a turn-level counterfactual information-gain reward: it measures how much a user’s feedback changed what the agent would do next, and uses that signal (combined with outcome rewards) to train agents on when to ask. Across UserGym, ColBench, and τ²-Bench, it outperforms both prompting and multi-turn RL baselines.

(Source: Atoms — ICML 2026 Papers Review)

AOrchestra: Dynamic Sub-Agent Creation

Most multi-agent systems use manually predefined roles. AOrchestra models every sub-agent as a four-tuple — instruction, context, tools, model — and lets a central orchestrator create specialized agents on demand per subtask. Across GAIA, Terminal-Bench 2.0, and SWE-Bench-Verified, it improves on static-role baselines. It’s also one of the first papers to show that orchestration can be optimized through supervised fine-tuning and cost-aware in-context learning.

AutoWebWorld: Synthetic Verifiable Web Environments

Training web agents requires interaction trajectories that are expensive to collect and hard to verify. AutoWebWorld generates web environments from finite state machine specifications, then uses breadth-first search over the known transition graph to produce verified trajectories — at $0.04 per trajectory. The result: 29 diverse environments, 11,663 verified trajectories, and consistent performance gains on WebVoyager and Online-Mind2Web when training on synthetic data.

MindFlow: Structured Thinking Flows for Research Ideation

Research ideation is open-ended and multi-objective. MindFlow models it as a graph over thinking operators — divergent, convergent, critical, analogical, counterfactual, constraint-driven — sampled from a probabilistic mind supernet and optimized through tournament-based relative ranking. The result is an ideation system that adapts its reasoning strategy per topic, improving novelty, diversity, and feasibility across domains.

InteractComp: Evaluating Search Agents Under Ambiguity

Search agents assume user queries are complete. InteractComp proves otherwise: 210 expert-curated ambiguous questions across 9 domains. The best model achieves 13.73% accuracy under ambiguity — versus 71.50% when given complete context. The bottleneck isn’t capability; it’s that agents are overconfident and don’t ask clarifying questions. Forced interaction dramatically improves performance, suggesting latent interactive capability that default strategies fail to activate.

Why This Matters for Production

The gap between ICML’s research energy and enterprise reality is the defining tension of the moment.

The global agentic AI market sits at $9.1–$10.9 billion in 2026, up from ~$7.3B in 2025, with analysts projecting 40%+ CAGR through decade’s end. Gartner forecasts 40% of enterprise applications will embed task-specific agents by end of 2026, up from under 5% in 2025. The intent is large-scale.

But the execution picture is more sobering. McKinsey puts scaled deployment near 23%. Gartner expects more than 40% of agentic AI projects will be cancelled by 2027, mostly from unclear value and inadequate risk controls. The gap between a demo that works and an agent that runs reliably in production is where ICML’s research is aimed.

(Sources: McKinsey — State of AI 2026, Gartner — Agentic AI Forecast)

Earlier this year, The Agent Report covered Anthropic’s Fable 5 deployment — a model so capable it triggered US export controls. The FAGEN workshop exists because capability is no longer the bottleneck; reliability is. When an agent can book flights, transfer money, or deploy code, the question isn’t whether it can act, but whether its failure modes are understood, reproducible, and fixable.

FAQ

Q: Why should builders care about an academic conference?

The techniques powering LLMs, RL agents, and generative systems — transformers, PPO, diffusion models — were first presented at ICML, NeurIPS, and ICLR. The papers accepted this week represent research that will shape production ML systems over the next 1–3 years. The agent safety and orchestration problems ICML is tackling now are the problems your deployment pipeline will hit next year.

Q: Is 23,918 submissions actually good for the field?

Mixed. It reflects a growing global research workforce and faster iteration cycles. But reviewers are stretched thin — the 26.6% acceptance rate means ~3,500+ qualified reviewers processed nearly 24,000 papers in a few months. Quality control at that scale is a real concern, and several prominent researchers have raised questions about review consistency.

Q: Why is agentic AI dominating ICML specifically?

Because the safety and reliability problems of autonomous agents sit at the intersection of the fields ICML has always owned: optimization, statistical learning theory, and reinforcement learning. When an agent takes irreversible actions in the real world, the question of whether its policy is safe stops being academic — and ICML’s theoretical toolkit is uniquely suited to answering it.

Q: What’s the most practical paper to read first?

If you deploy agents today: InteractComp (the ambiguity problem is universal). If you build multi-agent systems: AOrchestra (dynamic sub-agent creation replaces manual role design). If you’re an ML engineer: AutoWebWorld (synthetic training data at $0.04/trajectory).