RL algorithms
REINFORCE Zero: aligning execution before correcting policy lag
Why matching the learner and rollout engine changes what a simple asynchronous policy-gradient update can achieve.
Read the note →NVIDIA Research · Santa Clara, California
Hao Zhang张浩
Research Scientist working on agentic AI and reinforcement learning for capable, scalable language agents. Exploring multimodal intelligence, visual understanding, and efficient inference.
About
Building capable agents.
Understanding how they learn.
I’m a Research Scientist at NVIDIA Research, focusing on agentic AI and reinforcement learning for capable, scalable LLM agents. My research also spans large multimodal models, visual understanding, and efficient generative inference. I graduated in August 2024 from the Department of Computer Science & Engineering, Hong Kong University of Science and Technology, with a Ph.D. co-supervised by Prof. Heung-Yeung Shum and Prof. Lionel M. Ni. Previously, I obtained my bachelor’s degree from the School of Electronic Information and Electrical Engineering at Shanghai Jiao Tong University in 2019.
I welcome discussions and collaborations. Feel free to get in touch.
Research Scientist · NVIDIA Research
Ph.D. · HKUST
Research focus
Learning, acting, and scaling.
01
Agentic reinforcement learning
Training systems and algorithms for language agents that reason, use tools, and interact over long horizons.
02
Multimodal intelligence
Connecting language with images, video, and visual understanding to build agents that perceive and act.
03
Efficient inference
Making generative models and agent systems more practical through efficient decoding and model collaboration.
From the blog
Ideas behind the research.
Agent models
Nemotron 3.5: treating agent capability as a complete system
A look at how architecture, post-training, agent harnesses, and inference efficiency fit together in Super and Lightning.
Read the note →Agent models
AgentBaton: knowing when to hand the task over
Model collaboration becomes more interesting when the decision happens inside a task, after useful progress has been made.
Read the note →Research
Selected works.
-
REINFORCE Zero: Kernel Alignment Unlocks Stable Asynchronous RL with Importance Sampling
Technical report, 2026.
-
Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI
arXiv 2026.
-
AgentBaton: Efficient Model Collaboration in Agentic Tasks
arXiv 2026.
-
LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches
arXiv 2026.
-
Skill2Env: Reinforcing Agents with Collective Skills
Technical report, 2026.
-
OSWorld-Pro: Process-based Evaluation for Computer Use Agents
arXiv 2026.
-
Agora: Git as Shared Memory for Collective AutoResearch
NeurIPS 2026 AutoMLR Workshop. Oral presentation. arXiv 2026.
-
FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models
Technical report, 2026.
-
Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning
arXiv 2026.
-
Polar: Agentic RL on Any Harness at Scale
arXiv 2026.
-
ProRL Agent: Rollout-as-a-Service for RL Training of Multi-Turn LLM Agents
arXiv 2026.
-
Fast-dLLM v2: Efficient Block-Diffusion LLM
ICLR 2026.
-
Fast-dLLM: Training-Free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding
ICLR 2026.
-
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
arXiv 2025.
-
LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models
ICLR 2025.
-
LLaVA-OneVision: Easy Visual Task Transfer
arXiv 2024.
Earlier high-impact work
-
Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection
ECCV 2024.
-
DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection
ICLR 2023.
-
DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR
ICLR 2022.
-
DN-DETR: Accelerate DETR Training by Introducing Query DeNoising
CVPR 2022. Oral presentation.
-
SEEM: Segment Everything Everywhere All at Once
NeurIPS 2023.
-
Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation
CVPR 2023.
* denotes equal contribution.
Updates
News & milestones.
REINFORCE Zero, on kernel alignment for stable asynchronous RL with importance sampling, is released.
Nemotron 3.5 Super and Lightning, open hybrid models for agentic AI, is released.
AgentBaton and LoGRA, on efficient agent collaboration and memory-efficient RL, are released.
Skill2Env and OSWorld-Pro, on training and evaluating agents, are released.
Agora and FlashREINFORCE, on collective AutoResearch and asynchronous agentic RL, are released.
Molt, a scalable PyTorch-native training framework for agentic reinforcement learning, is released.
Earlier updates
Polar, a scalable rollout framework for agentic RL on arbitrary harnesses, is released.
Nemotron 3 Nano Omni, an efficient open multimodal model, is released.
ProRL Agent, rollout-as-a-service for multi-turn LLM-agent RL training, is released and integrated with NVIDIA NeMo Gym.
Fast-dLLM and Fast-dLLM v2 are accepted to ICLR 2026.
GR00T N1, an open foundation model for generalist humanoid robots, is released.
LLaVA-NeXT-Interleave is accepted to ICLR 2025.
Grounding DINO, a highly cited open-set object detector, is presented at ECCV 2024.
Mask DINO, a unified framework for object detection and segmentation, is presented at CVPR 2023.
DINO, an end-to-end detector with improved denoising anchor boxes, is presented at ICLR 2023.
DN-DETR is presented as a CVPR 2022 oral.
Recognition
Selected awards.
- CSE Best PhD Dissertation Award Honorable Mention, 2023–24
- RedBird Research Scholarship, HKUST, 2020
- Hong Kong Postgraduate Scholarship, 2020, 2021, 2022 & 2023