Hooray~ you caught me! I am Junhao Hu, a fifth-year Ph.D. student at School of Computer Science, Peking University, advised by Prof. Tao Xie. And I am lucky to work with and advised by Prof. Chenxi Wang, Prof. Yizhou Shan, Hailin Zhang, Kechi Zhang, and Xusheng Chen.

I am currently an RSI researcher at Tencent Hunyuan, working on RSI for harness and infrastructure. Feel free to reach out for discussion or collaboration!

I got my bachelor’s degree from Kuang Yaming Honors School, Nanjing University in 2022, where I did research on Compilation Resource Prediction, advised by Prof. Yanyan Jiang.

My research centers on algorithm–infrastructure co-design for LLMs and agents. I started with Infra4AI: optimizing inference infrastructure through serving systems, model–infra co-design, sparse attention, and position-independent caching (e.g., EPIC, DeepServe, RedKnot, Hy4, MiMo-V2). Then, as agents grew more powerful, I began building agent harnesses and memory systems (e.g., PenguinHarness, Akashic). Later, as AI grew more capable, building infrastructure by hand is no longer enough: rather than compete with AI, I train AI to build better infrastructure (AI4Infra). I now work on RSI for harness: training models to absorb harness capabilities such as compaction, skills, and self-evolution, and to write and evolve harnesses themselves (e.g., GDPevo, Awesome-RSI). Next, I plan to extend AI4Infra to inference infrastructure and even training infrastructure.

For more information, please hang around this website or refer to my CV.