Haoyang Hong

I am a Ph.D. student at Oregon State University, advised by Huazheng Wang. My research focuses on reinforcement learning theory, bandit algorithms, and statistical inference. I am interested in how learning algorithms behave when models are imperfect, rewards are unreliable, or data are collected adaptively.

I received my B.S. in Statistics from the University of Science and Technology of China (USTC), in the School of the Gifted Young. From July to October 2026, I was a research intern at Microsoft Research Asia, working with Wei Chen on reward hacking and adaptive reward proxy refinement.

Research

My work connects algorithm design with statistical analysis. Three questions guide my research.

• How do modeling assumptions affect reinforcement learning? I study KL regularization and function approximation, with a focus on what changes when the model class cannot fully represent the environment.
• When does optimizing a reward lead to the wrong outcome? I study reward poisoning and reward hacking, including how additional feedback can help identify and refine unreliable reward proxies.
• How should a learner collect information? I study bandit problems that balance decision quality, evaluation cost, and statistical inference, including improving proxies and network interference.

Publications and Preprints

* Equal contribution. See also Google Scholar.

@inproceedings{hong2026onlinekl,
  title={Online KL-Regularized Reinforcement Learning with Function Approximation under Misspecification},
  author={Hong, Haoyang and Wang, Zichen and Gu, Quanquan and Wang, Huazheng},
  booktitle={Reinforcement Learning Conference},
  year={2026}
}

@inproceedings{aguilarescamilla2026poison,
  title={When Can You Poison Rewards? A Tight Characterization of Reward Poisoning in Linear MDPs},
  author={Aguilar Escamilla, Jose Efraim and Hong, Haoyang and Li, Jiawei and Zhao, Haoyu and Zhang, Xuezhou and Hong, Sanghyun and Wang, Huazheng},
  booktitle={International Conference on Machine Learning},
  year={2026}
}

@article{lu2026beyond,
  title={Beyond Static Bias: Adaptive Multi-Fidelity Bandits with Improving Proxies},
  author={Lu, Muyun and Hong, Haoyang and Wang, Huazheng and Lin, Ying},
  journal={arXiv preprint arXiv:2605.08558},
  year={2026}
}

@inproceedings{wang2025design,
  title={Design-Based Bandits under Network Interference: A Trade-off between Regret and Inference},
  author={Wang, Zichen and Hong, Haoyang and Li, Chuanhao and Li, Haoxuan and Zhang, Zhiheng and Wang, Huazheng},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025},
  url={https://arxiv.org/abs/2510.07646}
}

Research Experience

Research Intern · Mentor: Wei Chen

My internship explored a practical question in reward hacking. How can occasional, costly feedback improve a cheap reward proxy, and when should the system request further evaluation?

• Investigated extensions of existing reward hacking theory to understand how optimizing a proxy reward affects the true reward and how additional feedback can guide proxy refinement.
• Studied adaptive methods for refining reward proxies using feedback from more capable LLMs, human supervision, or comprehensive environment evaluations. The aim was to reduce reliance on costly feedback as training progresses, while flagging potential proxy failures for further evaluation.
• Conducted preliminary experiments with coding agents and robotics tasks to examine proxy refinement, discrepancies between proxy and true rewards, and evaluation cost.

Education

Contact

I welcome conversations about reinforcement learning theory, bandits, and reward robustness.

honghao@oregonstate.edu
Oregon State University · Corvallis, Oregon, USA