Huan-ang Gao (高焕昂)

Ph.D. student, Tsinghua University

Mail: cc7w [at] foxmail [dot] com

Photo of Huan-ang Gao

Education

  • Tsinghua University logo
    Ph.D. student in Computer Science 2024 - Present
    Tsinghua University

    Advisor: Prof. Ya-Qin Zhang, Founding Dean of Institute for AI Industry Research (AIR), Tsinghua University.

  • Tsinghua University logo
    B.Eng. in Computer Science 2020 - 2024
    Tsinghua University

    GPA 3.98 / 4.00, ranked #1 out of 204 in the Department of Computer Science and Technology.

    Received the highest distinction for undergraduates at Tsinghua in 2023 and the SenseTime Scholarship in 2024.

Research Interests

I am a firm believer in The Bitter Lesson. What matters is general methods that scale with computation, and today that means scaling learning through agentic rollouts and environment feedback.

I am broadly interested in bringing LLM agents into every specialized field where real work gets done, as a path toward general-purpose agents. I think about this along three lines:

  1. Data & Evaluation. An RL environment consists of the instructions given to the agent, a sandbox equipped with the right tools, realistic data for the agent to work on, and a verifier that reliably judges the outcome. I care about deciding which tasks should be evaluated first, and then synthesizing them at scale.
  2. Scalable Rollouts. What if current models cannot produce meaningful learning signals in a new domain? I am particularly interested in self-evolving agents that accumulate reusable skills, and in mid-training: what kind of mid-training data effectively improves performance in a domain, and more broadly, what role data plays.
  3. Scalable Learning. How can we design learning algorithms that work efficiently for a given domain? My core belief is that learning algorithms are grounded in data: the data you have determines which algorithms you can use.

If you're interested in related topics and would like to collaborate, feel free to reach out!

Past Research

† first or co-first author

  • Agentic Reinforcement Learning

    I had a wonderful year-long research journey as an intern at ByteDance Seed, where I was a core contributor to the coding agent team for Seed 1.8 and Seed 2.0.

    During this time, I worked on both infrastructure and algorithms. On the infrastructure side, I contributed to building the RL training infrastructure and the data infrastructure behind it. On the algorithm side, I designed difficulty-aware data scheduling: when the data pool exceeds the training budget, the sampling ratio of each problem is adjusted dynamically according to its difficulty, together with mechanisms such as a replay buffer to reuse informative samples.

    • † Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation, arXiv 2026
    • † Weak-to-Strong Generalization via Direct On-Policy Distillation, arXiv 2026
    • CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation, arXiv 2026

  • Generative Simulation (Now termed as) World Models for Embodied AI

    The development and iteration of autonomous driving and robotics policies are limited by the high costs, low efficiency, and safety risks of real-world testing. Generative AI offers a potential breakthrough by enabling high-fidelity, interactive, and editable simulation testing.

    I worked on building world models that drive simulation with generative methods, along two directions: high-fidelity scene reconstruction, building digital twins of real scenes with NeRF or Gaussian splatting; and controllable content generation, leveraging the generative priors of diffusion models to provide endless, controllable scene variations and edge cases on top of reconstructed scenes.

    • † PartRM: Modeling Part-Level Dynamics with Large Cross-State Reconstruction Model, CVPR 2025
    • † Ctrl-U: Robust Conditional Image Generation via Uncertainty-aware Reward Modeling, ICLR 2025
    • † SCP-Diff: Spatial-Categorical Joint Prior for Diffusion Based Semantic Image Synthesis, ECCV 2024

  • Data Efficient Scene Parsing

    2D/3D perception is fundamental to embodied intelligence, but the extremely high cost of data annotation severely restricts the development of perception models.

    My early research focused on data-efficient perception, particularly semi-supervised learning and domain adaptation. In DQS3D, I proposed a single-stage, densely-matched semi-supervised framework for 3D object detection, addressing the insufficient training signals caused by sparse matching in previous methods. I also explored self-supervised depth estimation, indoor layout estimation, and HD map generation.

    • † DQS3D: Densely-matched Quantization-aware Semi-supervised 3D Detection, ICCV 2023
    • † From Semi-supervised to Omni-supervised Room Layout Estimation Using Point Clouds, ICRA 2023
    • † Training-Free Model Merging for Multi-target Domain Adaptation, ECCV 2024

Services

  • Co-Founder @ Lumina-Embodied.AI Research Community

    Building a community for embodied AI research and applications.

  • Reviewer

    Computer Vision: CVPR, ICCV, ECCV, WACV, 3DV

    Machine Learning & AI: NeurIPS, ICML, ICLR, AAAI

    Robotics: CoRL, ICRA, IROS

    Journal: TPAMI

  • Teaching Assistant @ Department of Computer Science and Technology, THU

    Software Engineering (23Spring - 25Fall)

    Digital Logic Experimentation (24Spring, 25Spring)

    Computer Organization and Design (23Fall)

  • 清华大学计算机科学与技术系 科创辅导员 (2024.9 - 2026.6)

    Technical training, competition guidance, and research & internship opportunities for undergraduates.

  • 清华大学计算机科学与技术系 学生科协主席 (2023.5 - 2024.6)

    Infrastructure: built the association's homepage and technical documentation.

    Training: organized the summer training camp for freshman year students.

    Resources: launched an internal resource portal gathering internship opportunities and graduate admissions information for students in the department.