Yearly, the Worldwide Convention on Machine Studying (ICML) reveals the place hundreds of AI researchers have determined to place their work.
This yr’s accepted papers reveal a transparent path: open frontier fashions and open AI infrastructure have turn out to be foundational to how fashionable AI science will get performed.
NVIDIA had 74 papers accepted at ICML 2026. Roughly 2,000 accepted papers cite NVIDIA GPUs, and 145 cite NVIDIA Nemotron — a household of open fashions, together with open datasets — as the inspiration for brand spanking new analysis. Tons of extra draw on NVIDIA Cosmos, NVIDIA Isaac GR00T, BioNeMo and different NVIDIA open mannequin households, spanning bodily AI, robotics, autonomous autos and biomedical analysis.
The Themes Defining This 12 months’s Analysis
Areas together with imaginative and prescient and video technology, reinforcement studying for big language fashions (LLMs) and agent coaching in addition to AI inference remained distinguished themes throughout this yr’s papers, reflecting sustained funding these fields command — whereas a number of new areas additionally broke via.
Robotic world fashions drew vital consideration, with papers like DreamDojo pushing the boundary of how AI methods be taught to cause about and act in bodily environments. DreamDojo, for instance, learns how the bodily world behaves from human video and builds on NVIDIA Cosmos open frontier fashions to foretell how a robotic would deal with objects and function in environments it was by no means skilled on. It lets researchers consider insurance policies, plan actions and teleoperate a digital robotic, accelerating growth with out the prices and dangers of bodily deployment.
AI for all times sciences was fueled by NVIDIA BioNeMo open fashions and analysis contributions that assist researchers perceive protein operate, molecular habits and genetic code. Papers like FLIP2 introduce public benchmarks for testing how properly AI predicts the results of protein mutations. KERMT is a brand new BioNeMo open mannequin for predicting molecular properties essential to drug discovery.
Artificial information technology (SDG) drew specific curiosity at ICML this yr with a number of Nemotron and bodily AI open datasets, reflecting a broader shift in how researchers are fascinated by coaching at scale with out relying solely on human-labeled information.
The Open Analysis Stack
Open infrastructure provides researchers the instruments to speed up breakthroughs.
The papers present Nemotron getting used much less like a single mannequin launch and extra like a analysis stack: open weights to judge towards, open datasets to coach and adapt with, and open recipes for reasoning, device use, security, information curation and environment friendly inference.
Alongside the fashions, NeMo Curator and the open datasets it helps provides researchers a reproducible basis for coaching information curation. SDG instruments allow creating high-quality coaching units at a scale and velocity that may’ve been impractical only a few years in the past.
The Cosmos 3 household of open, frontier omnimodels provides researchers and builders a generational leap within the skill to construct robots, autonomous autos and imaginative and prescient AI that understand, cause, plan and act within the bodily world.
As well as, the NVIDIA Alpamayo open mannequin household for autonomous car growth, NVIDIA Isaac GR00T for robotics and NVIDIA BioNeMo for biomedical assist speed up analysis and growth throughout industries.
The Ecosystem Constructing on High
The momentum extends past NVIDIA’s personal analysis labs.
Basecamp Analysis developed a brand new DNA basis mannequin, EDEN, that helps researchers interpret and design genetic sequences.
Merck & Co., makes use of KERMT to foretell how potential drug molecules might behave within the physique, together with whether or not they’re prone to be efficient, secure and developable.
Sakana AI — attending ICML this yr — constructed its Fugu and Fugu-Extremely fashions straight on Nemotron 3 Extremely, utilizing the open basis to push ahead its work on AI analysis automation.
KiloCode built-in Nemotron into its code-routing structure, reporting token price reductions of as much as 90% — a end result with actual implications for the economics of deploying AI in manufacturing.
NAVER developed its personal mannequin utilizing the Nemotron structure, extending the inspiration for Korean-language AI analysis.
Collectively AI is internet hosting Nemotron fashions on its platform, making them extra accessible to researchers who want dependable, seamless entry to open inference.
Humanoid, LG Electronics, NEURA Robotics and Noble Machines are adopting NVIDIA Isaac GR00T fashions to speed up industrial deployments of their humanoids, whereas 1X, Agility, Agile Robots, Boston Dynamics, Hexagon Robotics, and Mentee are constructing the subsequent technology of humanoids utilizing Cosmos world fashions, Isaac Sim and Isaac Lab to speed up the event and validation of their robots.
Discover NVIDIA’s open fashions on Hugging Face.
Discover genomics and biology analysis at ICML’s GenBio Workshop on Friday, July 10.
