Each year, the International Conference on Machine Learning (ICML) charts the course for thousands of AI researchers, indicating their research priorities.
This year’s accepted papers highlight a pronounced focus on frontier models and AI infrastructure, serving as the foundation for contemporary AI science.
NVIDIA proudly accepted 74 papers at ICML 2026, with around 2,000 papers referencing NVIDIA GPUs and 145 mentioning NVIDIA Nemotron. This includes a collection of open models and an open dataset that form the basis for innovative research. Numerous features are also available through NVIDIA Cosmos, Isaac GR00T, and BioNeMo, promoting advancements in fields like robotics, self-driving vehicles, and biomedical research.
Themes Defining This Year’s Research
Key areas include vision and video generation and reinforcement learning, particularly concerning large language models (LLMs) and agent training. The field of AI inference also remains a significant theme, reflecting consistent investments in these sectors, with new fields emerging to keep pace.
Robotics and world models garnered notable interest, particularly with works like DreamDojo. This model extends our understanding of how AI systems can learn and function in physical settings. For instance, DreamDojo derives insights about the physical world from human videos and leverages the NVIDIA Cosmos Open Frontier Model to predict robotic interactions within unfamiliar environments, facilitating policy assessments and action planning while minimizing the risks linked to physical experiments.
AI in Life Sciences is driven by NVIDIA’s BioNeMo open models, aiding scientists in deciphering protein functions, molecular behaviors, and genetic codes. A notable paper, Flip 2, proposes a public benchmark for evaluating how AI predicts protein mutation effects. The Celtic model is a recent BioNeMo endeavor aimed at predicting essential molecular properties for drug discovery.
Synthetic Data Generation (SDG) attracted significant interest at ICML, showcasing several Nemotron and Physics AI projects, reflecting a shift towards training at scale without becoming overly reliant on human annotations.
Open Research Stack
Open infrastructures empower researchers with the tools needed to accelerate breakthroughs in AI.
Recent findings illustrate that Nemotron is evolving into a comprehensive research stack rather than just a standalone model. This includes open weights for evaluation, open datasets for training and adaptation, and open methodologies for inference and safe tool utilization.
Beyond models, NeMo Curator offers open datasets that provide a reproducible framework for data curation essential for training, with SDG tools that deliver high-quality training set creation at unprecedented rates.
Through Cosmos 3, the open frontier model family is equipping researchers and developers to achieve unprecedented advancements in robotics, autonomous vehicles, and AI-based vision technologies that can perceive, reason, plan, and act within their environments.
Additionally, NVIDIA Alpamayo is contributing to the development of autonomous vehicles, while NVIDIA Isaac GR00T bolsters robotics and NVIDIA BioNeMo aids biomedical science, accelerating research across various sectors.
Building the Ecosystem
The momentum extends beyond NVIDIA alone, influencing research initiatives in various laboratories. NVIDIA’s labs are fostering innovation.
Base Camp Research is designing a pioneering DNA model, EDEN, which supports scientists in interpreting and designing gene sequences.
Merck & Co. employs Celtic to anticipate the behavior of drug molecules within biological contexts, evaluating their efficacy and safety.
Fish AI, actively participating in ICML this year, is leveraging Fugu Fugu-Ultra to construct models grounded in Nemotron 3 Ultra, harnessing open infrastructure to enhance AI research automation.
KM Code has integrated Nemotron into its code routing framework, achieving token cost reductions of up to 90%, a significant milestone for AI deployment economics.
Naver developed its own model using the Nemotron architecture, enriching the foundation of Korean AI research.
AI Together hosts Nemotron models on its platform, facilitating seamless access to open inference for researchers.
Companies like Humanoid, LG Electronics, Neural Robotics, and Noble Machine are implementing the NVIDIA Isaac GR00T model to enhance humanoid applications. Collaborating alongside Boston Dynamics, Hexagon Robotics, and Mentee, they are accelerating the development and validation of next-generation humanoid robots using the Cosmos world model, Isaac Sim, and Isaac Lab.
Explore NVIDIA’s open model offerings via Hugging Face.
Join the exploration of genomics and biology at ICML during the GenBio Workshop on Friday, July 10th.
Source: blogs.nvidia.com