Han (Paris) Zhang 章晗

I am an Applied Research Scientist at NVIDIA, where I build VLMs that perceive, reason, and act on real-world video for Physical AI and Visual AI Agents, and have contributed to the Cosmos and Nemotron VL series.

My work spans the full stack of VLM post-training, including agentic data engines, SFT and RL recipes, benchmarks, and real-world deployment. I lead post-training of the Cosmos VLMs for Metropolis, authored the fine-tuning recipes in NVIDIA TAO and Cosmos Cookbook, and co-organized AI City Challenges.

I earned my Master's degree in Computer Science from Stanford University, where I was honored as a Siebel Scholar, Class of 2025.

I received my B.A. in Computer Science and Economics from UC Berkeley, where I was awarded the prestigious EECS Department Citation. During my time there, I had the privilege of conducting research in vision and language with Prof. Trevor Darrell, Prof. Joseph E. Gonzalez, and Ph.D. candidate Lisa Dunlap.

Email  /  Google Scholar  /  LinkedIn

My research interests broadly span computer vision, with a particular focus on vision-and-language multimodalities, generative models, and the robustness and interpretability of AI systems.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. I contributed the Smart Infrastructure data and Traffic Anomaly Reasoning (TAR) benchmark, and integrated the intelligent-transportation reasoning capability and data mixture into the Reasoner model.

We present MAVEN, a multi-stage agentic pipeline that turns raw videos into multi-task training data with CoT reasoning traces, organized around a designated Event of Focus. Agentically, MAVEN supports agent-driven domain adaptation and a hierarchical refinement loop.

We present CHURRO, a 3B-parameter open-weight VLM specialized for historical text recognition. The model is trained on CHURRO-DS, the largest historical text recognition dataset to date.

We introduce ALIA, a method which utilizes large vision and language models to automatically generate natural language descriptions of a dataset's domains and augment the training data via language-guided image editing.

We propose LADS, a method that learns a transformation of the image embeddings from the training domain to each unseen test domain guided by language, while preserving task relevant information.

Leading the post-training (SFT + RL) of Cosmos VLMs in Metropolis for video understanding and reasoning.

Trained and released Commercial VILA and built VLM auto-labeling pipelines generating 3M+ SFT samples that contributed to the data mixture of VILA and Nemotron Nano VL series. Developed a 3D-VLM for warehouse spatial reasoning featured at GTC 2025. Filed multiple patents on VLM data auto-labeling.

Developed and modularized a calculator for the loan product, GS Select, enabling automated daily computation of key financial metrics, including risk-weighted assets.

Optimized TikTok's Android package size and cold startup time by developing custom optimization passes leveraging an open-source Android bytecode optimizer, ReDex.

Website template from Jon Barron. Thanks for stopping by :)
Last updated: July 8, 2026