Long-Horizon Agent Design
Code agents that reliably finish long-horizon tasks: scaling the harness around the model, designing robust agent loops, and training agents with cost-aware RL.
01About
I am a Principal Researcher at Tencent, leading efforts on code agents for long-horizon tasks. I also collaborate closely with Prof. Atlas Wang.
I received my Ph.D. in 2025 from the NUS HPC-AI Lab, advised by Prof. Yang You. During my Ph.D., I was an affiliated member at the University of Toronto with Prof. Konstantinos N. Plataniotis, and a sponsored researcher at Carnegie Mellon University with Prof. Alex Hauptmann. Before that, I received my M.S. in 2020 from MMLab@SIAT, advised by Prof. Yu Qiao and Assoc. Prof. Xiaojiang Peng.
I like simple, strong baselines that reveal how deep learning actually works. Representative work: HY-WU, DnD (NeurIPS 2025), RPG (NeurIPS 2025), p-diff, and PAB (ICLR 2025).
02News
03Research
Parameter Generation
Treating neural network weights as a generative modality: generating, adapting, and composing parameters directly.
Efficient Generation & Training
Making model training and diffusion/video generation faster and cheaper, often with simple tricks that drop into existing pipelines.
Data-Centric AI
Getting more out of less data: dataset distillation, dataset quantization, and dynamic data pruning for lossless, faster training.
04Publications
Full list on Google Scholar* equal contribution · + corresponding author
05Experience & Education
06Mentoring
I've been lucky to work with these talented students.