Photo of Shu Zhao

Shu Zhao

Senior Machine Learning Engineer

NVIDIA/Santa Clara, CA

Curiosity drives me to keep building on what I know and connecting ideas across fields. My research began in computer vision during my master's, grew to include multimodal learning during my Ph.D., and now focuses on LLM agents at NVIDIA. At the moment, I'm particularly interested in how post-training can turn LLMs into agents that get real work done.

Outside of work, I love driving and taking in the scenery. My curiosity turned that into a side project: Joi (named after Blade Runner 2049), an AI co-pilot that runs on a Jetson Nano in my car, gives me road and safety information, and chats with me along the way.

I don't know where my curiosity will take me next, and that's part of the fun.

News

  1. Our new paper, Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation, is on arXiv. We also released miles-agent, a post-training framework for research on long-horizon agents.

  2. The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors was accepted at NeurIPS 2026 as a Spotlight.

  3. Joined NVIDIA as a Senior Machine Learning Engineer.

  4. MoRA: Missing Modality Low-Rank Adaptation for Visual Recognition was accepted at ICLR 2026.

  5. Completed my Ph.D. at Penn State.

Earlier

  1. Released ferret, an extensible RL framework for training LLM agents with advanced search capabilities.

  2. HRScene: How Far Are VLMs from Effective High-Resolution Image Understanding? was accepted at ICCV 2025.

  3. SafeMap: Robust HD Map Construction from Incomplete Observations was accepted at ICML 2025.

  4. I will be interning at NVIDIA this summer. See you in Santa Clara!

Selected Research

Post-training for long-horizon agents

Multimodal learning

Efficient image retrievalWhere my research began

Open Source

miles-agent Post-training for long-horizon agents: RL and on-policy distillation on miles, with NeMo Gym environments and Harbor sandboxes for terminal and coding rollouts. GitHub ferret An extensible RL framework for search agents on verl: modular rewards and tools, with end-to-end recipes for ParallelSearch, ExpandSearch, and Search-R1. GitHub

Experience & Education

Work

  1. 2026 – nowSenior Machine Learning EngineerNVIDIA
  2. 2025Machine Learning Engineer InternNVIDIA
  3. 2022Research InternBaidu Research

Education

  1. 2022 – 2025Ph.D., Computer ScienceThe Pennsylvania State University
  2. 2018 – 2021M.E., Computer TechnologyUniversity of Chinese Academy of Sciences
  3. 2013 – 2017B.E., Information SecurityAnhui University