ICML 2026 Papers Show Open AI Models Driving Research Momentum

The International Conference on Machine Learning has become a reliable indicator of where AI research is heading. This year's accepted papers signal a decisive shift: open frontier models and accessible AI infrastructure now form the backbone of cutting-edge research across robotics, life sciences, and synthetic data generation.

NVIDIA reported 74 accepted papers at ICML 2026, but the broader pattern is more revealing. Approximately 2,000 accepted papers cite NVIDIA GPUs in their work, while 145 specifically reference Nemotron—an open model family that includes datasets and training recipes. Hundreds more draw on Cosmos for physical AI, Isaac GR00T for robotics, and BioNeMo for biomedical applications, according to details first reported by NVIDIA.

Why it matters

The research community's adoption of open models represents a structural change in how AI science advances. When thousands of researchers build on shared foundations rather than proprietary systems, breakthroughs compound faster and reproducibility improves. For enterprises, this means the AI tools entering production are increasingly validated by a global research community rather than developed behind closed doors.

Research themes emerging at ICML

Robot world models attracted significant attention this year. DreamDojo, one notable paper, demonstrates how AI systems can learn physical reasoning from human video footage. Built on NVIDIA Cosmos open models, it enables robots to predict object handling and environmental navigation without prior training in those specific scenarios. Researchers can evaluate policies and plan actions in virtual environments before risking physical deployment.

Life sciences research leveraged BioNeMo open models to advance understanding of protein function and molecular behavior. FLIP2 introduces public benchmarks for predicting protein mutation effects, while KERMT—a new BioNeMo model—focuses on molecular properties relevant to drug discovery.

Synthetic data generation emerged as a particularly active area, with multiple papers exploring Nemotron and physical AI datasets. This reflects growing recognition that training at scale cannot rely exclusively on human-labeled data.

The open research stack in practice

Researchers are treating Nemotron less as a single model and more as a complete research infrastructure. It provides open weights for evaluation, datasets for training and adaptation, and documented approaches for reasoning, tool use, safety protocols, and data curation.

The Cosmos 3 family delivers what NVIDIA describes as a generational advancement in building robots and autonomous vehicles that can perceive and act in physical environments. Meanwhile, NeMo Curator offers reproducible methods for training data preparation.

Industry adoption patterns

The momentum extends beyond academic labs. Basecamp Research developed EDEN, a DNA foundation model for genetic sequence interpretation. Merck uses KERMT to predict drug molecule behavior, including efficacy and safety profiles.

Sakana AI built its Fugu and Fugu-Ultra models directly on Nemotron 3 Ultra for AI research automation. KiloCode integrated Nemotron into code-routing systems and reported token cost reductions reaching 90 percent—a result with direct implications for production deployment economics.

NAVER extended the Nemotron architecture for Korean-language research, while Together AI now hosts Nemotron models to provide researchers with reliable inference access.

In robotics, companies including Humanoid, LG Electronics, NEURA Robotics, and Noble Machines are adopting Isaac GR00T models for industrial humanoid deployments. Others like 1X, Agility, Boston Dynamics, and Hexagon Robotics are using Cosmos world models with Isaac Sim and Isaac Lab for development and validation.

These details were first reported by NVIDIA in its ICML 2026 research summary.