NVIDIA research at ICML 2026: 145 papers cite Nemotron open models; 2,000 papers use NVIDIA GPUs

NVIDIA had 74 papers accepted at ICML 2026, with approximately 2,000 total accepted papers citing NVIDIA GPUs and 145 papers using Nemotron open models as their research foundation. Papers showcase DreamDojo (robot world models learning from human video), FLIP2 (protein mutation benchmarks), and KERMT (new BioNeMo open model for molecular property prediction in drug discovery). The concentration of ICML acceptances around NVIDIA infrastructure and open models reflects industrywide shift: researchers increasingly build on proprietary-weighted foundation models and open datasets rather than training dense, monolithic architectures from scratch.

Synthetic data generation (SDG) drew particular ICML attention, with Nemotron and physical AI open datasets used across multiple accepted papers. The research stack model—open weights, open datasets, open recipes for reasoning and tool use—emerged as the preferred baseline for new work. NeMo Curator and standardized dataset tools enable reproducible training data curation at scale. Papers combining robot world models with multimodal reasoning (e.g., DreamDojo on NVIDIA Cosmos) signal that physical AI simulation and agentic reasoning are converging as research priorities.

For AI researchers and model builders, the ICML 2026 acceptance pattern confirms that open foundation models and infrastructure have become table stakes. Proprietary training still matters, but the research community now measures productivity by how readily new ideas integrate with established stacks (NVIDIA Nemotron, Claude, Llama) rather than whether researchers build from atomic principles. This accelerates downstream adoption but raises questions about research diversity and the cost to enter the game at scale.