π§ About Me
I am a Principal Scientist at NVIDIA Fundamental Generative AI Research (GenAIR).
I lead the development of next-generation generative AI foundation models, with a focus on scaling principled algorithms from theory to real-world impact.
My work bridges theory and practice of generative learning to advance generative AI. I am deeply interested in both algorithmic/theoretical foundations of generative modeling, and its real-world applications across language, vision, and science.
I am currently focused on LLMs, diffusion language models, and training/inference strategies for fast agentic AI. Recent efforts include scaling continuous latent diffusion for text, advancing tri-mode language models that unify autoregressive and diffusion decoding, and improving high-throughput generation.
Previously, I led major work on diffusion foundation models for scientific simulation, including core technology for NVIDIA’s Earth-2 digital twin. This work has been integrated into NVIDIA products and adopted by major partners such as The Weather Channel.
π Bio
I earned my PhD in Electrical Engineering with a minor in Mathematics from the University of Minnesota, advised by Georgios Giannakis, focusing on statistical learning.
I was a Visiting Scholar at UC Berkeleyβs RISE Lab with Michael Mahoney, and later a Postdoc and Research Associate at Stanford, working with David Donoho, John Pauly, and Shreyas Vasanawala on inverse problems and generative models.
I serve as an IEEE SPS Distinguished Industry Speaker, and a recipient of the IEEE SPS Young Author Best Paper Award.
π° News
- 2026 β Released Nemotron-Labs-Diffusion: the first tri-mode language model family that unifies autoregressive, diffusion, and self-speculation decoding in one set of weights (switchable via attention masks at inference), delivering strong quality with up to ~4x single-user throughput gains. Tech report | Open-source models
- May 18, 2026 β Posted on arXiv: Continuous Diffusion Scales Competitively with Discrete Diffusion for Language (arXiv)
- April 3, 2026 β Posted on arXiv: Understanding Latent Diffusability via Fisher Geometry (arXiv)
- March 7, 2026 β Posted on arXiv: Variational Flow Maps: Make Some Noise for One-Step Conditional Generation; accepted at ICML 2026 (arXiv)
- January 26, 2026 β Released the ATLAS foundation model and posted: Demystifying Data-Driven Probabilistic Medium-Range Weather Forecasting (arXiv)
- February 27, 2026 β Invited talk at UC Irvine: Steering Diffusion Models for Next-Gen AI
- October 2025 β Keynote at DigiAdvance (Aveiro, Portugal) and invited talk at Instituto Superior TΓ©cnico (IST), Lisbon: Steering Diffusion Models for Next-Gen AI
- September 2025 β NeurIPS 2025 paper accepted: Test-Time Scaling of Diffusion Models via Noise Trajectory Search (arXiv)
- September 2025 β NeurIPS 2025 paper accepted: Elucidated Rolling Diffusion Models (arXiv)
- January 2025 β ICLR 2025 paper accepted: Heavy-Tailed Diffusion Models (arXiv)
- September 2024 β Elected as IEEE Distinguished Industry Speaker
πΌ Opportunities
I am actively looking for PhD interns and collaborators interested in advancing generative modeling.
If you’re excited about building principled and impactful generative models, get in touch:
π§ mmardani@nvidia.com