Every year, the International Conference on Machine Learning (ICML) showcases where thousands of AI researchers have decided to place their work.
This year’s accepted contributions show a clear direction: open Boundary models and open AI infrastructure have become the foundation for the way modern AI science is conducted.
74 entries from NVIDIA were accepted at ICML 2026. Approximately 2,000 accepted posts cite NVIDIA GPUs and 145 cite NVIDIA Nemotron – a family of open models, including open datasets – as a basis for new research. Hundreds more use NVIDIA cosmosNVIDIA Isaac GR00T, BioNeMo and other open NVIDIA model families that include physical AI, robotics, autonomous vehicles and biomedical research.
Table of Contents
The topics that determine this year’s research
Areas such as vision and video generation, Reinforcement learning for large language models (LLMs) And agent training as well AI conclusion remained an important theme in this year’s contributions, reflecting continued investment in these areas – while several new areas also made breakthroughs.
robot World models attracted a lot of attention with articles like DreamDojo We are pushing the boundaries of how AI systems learn to think about and act in physical environments. DreamDojo, for example, learns how the physical world behaves from human videos and builds on NVIDIA Cosmos’ Open Frontier models to predict how a robot would handle objects and operate in environments in which it was never trained. This allows researchers to evaluate policies, plan interventions, and teleoperate a virtual robot, accelerating development without incurring the costs and risks of physical deployment.
AI for life sciences was driven by open NVIDIA BioNeMo models and research contributions that help researchers understand protein function, molecular behavior and genetic code. Papers like FLIP2 Launch public benchmarks to test how well AI predicts the effects of protein mutations. KERMT is a new open BioNeMo model for predicting molecular properties important for drug development.
Synthetic data generation (SDG) attracted particular interest at ICML this year with several Nemotron and physical AI Open datasets that reflect a broader shift in the way researchers think about training at scale, without relying solely on human-labeled data.
The Open Research Stack
An open infrastructure gives researchers the tools to accelerate breakthroughs.
The papers show that Nemotron is used less like a single model release and more like a research stack: open weights for evaluation, open datasets for training and customization, and open recipes for reasoning, tool usage, security, data curation, and efficient reasoning.
In addition to the NeMo Curator models and the open data sets it supports provides researchers with a reproducible basis for curating training data. SDG tools enable the creation of high-quality training sets at a scale and speed that would have been impractical just a few years ago.
The Cosmos 3 Open Border Family Omni models enables researchers and developers a generational leap in the ability to build robots, autonomous vehicles and visionary AI that sense, reason, plan and act in the physical world.
Furthermore, the NVIDIA alphabet open model family for autonomous vehicle development, NVIDIA Isaac GR00T for robotics and NVIDIA BioNeMo for biomedical assistance to accelerate research and development across industries.
The ecosystem based on it
The momentum goes beyond that of NVIDIA Research laboratories.
Basecamp Research developed a new basic DNA model, EDEN, to help researchers interpret and design genetic sequences.
Merck & Co.uses KERMT to predict how potential drug molecules might behave in the body, including whether they are likely to be effective, safe and developable.
Section AI – attended ICML this year – built its Fugu and Fugu Ultra models directly on Nemotron 3 Ultra and used the open foundation to advance its work on automating AI research.
KiloCode integrated Nemotron into its code routing architecture and reported token cost reductions of up to 90% – a result with real implications for the economics of using AI in production.
NAVER developed its own model using the Nemotron architecture, expanding the basis for Korean-language AI research.
AI together hosts Nemotron models on its platform, making them more accessible to researchers who need reliable, seamless access to open inference.
Humanoid, LG Electronics, NEURA Robotics And Noble machines adopt NVIDIA Isaac GR00T models to accelerate the industrial use of their humanoids 1X, agility, Agile robots, Boston Dynamics, Hexagon roboticsAnd Mentee are building the next generation of humanoids using Cosmos world models, Isaac Sim and Isaac Lab to accelerate the development and validation of their robots.
Discover NVIDIA’s open models Hugging face.
Discover genomics and biology research at ICML GenBio workshop on Friday, July 10th.
https://blogs.nvidia.com/blog/open-models-icml-2026/
