ICML 2026 has offered the clearest signal yet that open models are no longer a fringe preference — they're the default infrastructure for serious AI research. Of this year's accepted papers, approximately 2,000 cite NVIDIA GPUs, 145 cite NVIDIA Nemotron, and hundreds more draw on Cosmos, Isaac GR00T, and BioNeMo. NVIDIA itself had 74 papers accepted at the conference.

The Themes Shaping 2026

Several research threads dominated this year's program:

  • Vision and video generation
  • Reinforcement learning for LLMs and agent training
  • AI inference optimization
  • Robot world models
  • Synthetic data generation (SDG)
  • AI for life sciences

Robot world models attracted particular attention. DreamDojo — built on NVIDIA Cosmos open frontier models — learns physical-world behavior from human video, enabling robots to handle objects and navigate environments they were never explicitly trained on. It supports policy evaluation, action planning, and virtual teleoperation, cutting both cost and risk compared to physical deployment.

Open Models as a Research Stack

What's notable about Nemotron at ICML this year isn't any single model release — it's how researchers are treating the entire family as a modular stack.

  • Open weights for benchmarking and evaluation
  • Open datasets for training and domain adaptation
  • Open recipes covering reasoning, tool use, safety, data curation, and efficient inference

NeMo Curator and its associated datasets give researchers a reproducible pipeline for training data curation. SDG tooling now makes it practical to generate high-quality training sets at a scale that would have been impractical just a few years ago.

The Cosmos 3 family extends this to physical AI — giving developers a foundation for robots, autonomous vehicles, and vision systems that can perceive, reason, and act in the real world.

Life Sciences Gets a Boost

NVIDIA BioNeMo open models are fueling a wave of biomedical research at ICML:

  • FLIP2 introduces public benchmarks for predicting the effects of protein mutations
  • KERMT is a new BioNeMo open model targeting molecular property prediction for drug discovery
  • Basecamp Research developed EDEN, a DNA foundation model for interpreting and designing genetic sequences
  • Merck & Co. is using KERMT to predict how drug candidates may behave in the body

The Ecosystem Building on Top

The ripple effects extend well beyond NVIDIA's own labs.

Sakana AI built its Fugu and Fugu-Ultra models directly on Nemotron 3 Ultra, using the open foundation to advance AI research automation.

KiloCode integrated Nemotron into its code-routing architecture and reported token cost reductions of up to 90% — a meaningful result for anyone thinking about the economics of production AI deployment.

NAVER extended the Nemotron architecture for Korean-language AI research. Together AI is hosting Nemotron models on its platform to simplify open inference access for researchers.

On the robotics side, the adoption is broad:

  • Humanoid, LG Electronics, NEURA Robotics, and Noble Machines are using Isaac GR00T models for industrial humanoid deployments
  • 1X, Agility, Agile Robots, Boston Dynamics, Hexagon Robotics, and Mentee are building next-generation humanoids using Cosmos world models, Isaac Sim, and Isaac Lab

The picture that emerges from ICML 2026 is less about any single breakthrough and more about a structural shift: open models have become the shared substrate on which the broader AI research community builds.