NVIDIA points to ICML results as open AI models expand beyond software into robotics and biology

NVIDIA used the results of the International Conference on Machine Learning, one of the industry’s most influential academic venues, to argue that open models are becoming the default infrastructure for AI research. In an update published July 6, the company said accepted papers at ICML 2026 show a strong shift toward open frontier models and open AI tooling as the foundation for how researchers test, adapt, and build new systems.

NVIDIA said it had 74 papers accepted at ICML 2026. It also pointed to citation activity in the accepted paper set, estimating that roughly 2,000 accepted papers cite NVIDIA GPUs and that 145 cite Nemotron, a family of open models including open datasets. In the same framing, NVIDIA said “hundreds more” draw on other NVIDIA open model families including Cosmos, Isaac GR00T, and BioNeMo, spanning robotics, autonomous vehicles, and biomedical research.

Beyond citations, NVIDIA highlighted research themes it said remained prominent in the accepted papers. These included vision and video generation, reinforcement learning for large language models, agent training, and AI inference. NVIDIA said several new areas also appeared, including what it called “robot world models,” which focus on how AI systems learn about physical environments to plan and act.

For robot world models, NVIDIA pointed to papers such as DreamDojo. NVIDIA said DreamDojo learns how the physical world behaves from human video, builds on NVIDIA Cosmos open frontier models, and aims to predict how a robot would handle objects and operate in environments it was not trained on. NVIDIA also said the approach is meant to let researchers evaluate policies, plan actions, and teleoperate a virtual robot, with the goal of accelerating development without the costs and risks of deploying directly in physical spaces.

In life sciences, NVIDIA said BioNeMo open models and research contributions were a driver of activity at ICML 2026. It cited FLIP2 as an example, saying the paper introduces public benchmarks for testing how well AI predicts the effects of protein mutations. NVIDIA also said KERMT is a new BioNeMo open model aimed at predicting molecular properties relevant to drug discovery.

A separate area drawing attention, NVIDIA said, was synthetic data generation, or SDG. SDG refers to creating training data automatically, rather than relying solely on human-labeled examples. NVIDIA said SDG appeared alongside Nemotron and physical AI open datasets in ICML accepted work, reflecting what it called a broader shift in how researchers plan training at scale.

NVIDIA’s larger argument was not simply that its models are being used, but that open models are evolving into a more complete “research stack.” The company said Nemotron is used less like a single model release and more like a set of components, including open weights for evaluation, open datasets for training and adaptation, and open “recipes” for tasks such as reasoning, tool use, safety, data curation, and efficient inference. NVIDIA also described NeMo Curator as part of this ecosystem, characterizing it as a way to make training-data curation reproducible. For SDG, NVIDIA said its tools can produce higher-quality training sets faster than before.

The company also tied its open model families to specific application domains. It described Cosmos 3 as an open, frontier omnimodel family intended to help researchers build capabilities for robots, autonomous vehicles, and vision AI that can perceive, reason, plan, and act in physical settings. NVIDIA also referenced Alpamayo for autonomous driving research, Isaac GR00T for robotics, and BioNeMo for biomedical work, positioning them as accelerators for R and D across industries.

NVIDIA added that the momentum extends beyond its own research labs, citing several outside organizations that it said are building on its open foundations. It mentioned Basecamp Research working on a DNA foundation model called EDEN, Merck & Co. using KERMT for predicting how drug molecules behave in the body, and Sakana AI building models called Fugu and Fugu-Ultra on top of Nemotron 3 Ultra to advance AI research automation. It also cited a reported token-cost reduction of up to 90% after integration by KiloCode, and said NAVER used Nemotron architecture for Korean-language AI research. On deployment and access, NVIDIA said Together AI hosts Nemotron models on its platform, and it listed industrial and robotics companies that it said are adopting Isaac GR00T and Isaac world-model approaches built around Cosmos, Isaac Sim, and Isaac Lab.

What NVIDIA did not provide in its post is detail on the full breakdown of the 74 accepted papers, the specific methodology used to tally the citation estimates, or any performance benchmarks that prove open models outperform closed alternatives in every setting. The post also does not spell out licensing terms, compute requirements, or safety limitations for each model family, elements that often matter to researchers deciding what to adopt. As a result, the takeaway is strongest for directional evidence on where academic work is concentrating and which building blocks researchers appear to reference.

What to watch next is whether the open “stack” NVIDIA describes becomes a measurable commercial advantage, such as faster iteration cycles for robotics and life-science pipelines, or broader adoption of open inference services. For the sector, ICML accepted work can act as an early indicator of where research funding and developer tooling may flow, especially in physical AI, agent training, and synthetic data approaches. NVIDIA’s next statements on open model releases and tooling updates, as well as follow-on citations in future conferences, are likely to be closely watched by both researchers and infrastructure providers.