Beijing, ChinaI work on open-world visual understanding and large-scale foundation models that connect language, vision, and action.
My current focus is world models and vision-language-action (VLA), with an emphasis on practical, deployable intelligence.
I lead Wenyuan Pavilion, an open-source community dedicated to Chinese-culture-centered language models.
I have pursued computer vision research with Yian Zhao and Chaoran Feng since 2023, and explored LLMs with Xuxin Cheng since 2024.
Currently
Qianlima Talent Program
I lead a PhD research intern team developing next-generation VLA architectures with Hefei Huang, Jianyu Zhang, and Zhiyuan Liang.
World models and VLA: representation learning, data, post-training, and reinforcement learning.
We welcome talented researchers to join us.
Research Intern · Longcat Interaction Team · Beijing
Worked on ToB agentic RL for Table/Excel question answering. Led knowledge-graph-based data synthesis and developed Hindsight-guided Knowledge Distillation to reduce distillation steps and improve accuracy.
Institute of Automation, Chinese Academy of Sciences · Research Intern · 2023.11 – 2024.08
School of Artificial Intelligence, SJTU · Research Intern · 2025.07 – 2025.09
Developed EyeVLA for language-guided rotation and zoom, combining VLMs, 500 real-world examples, and RL post-training for robotic active perception.
Research Intern · Applications of LLMs · Beijing
Designed Deep Research data synthesis, supporting BrowseComp SOTA; contributed to LabelGPT video understanding and structured annotation.
Research Assistant
Multimodal document understanding with Prof. Huchuan Lu and Prof. Xu Jia.
An OS-like exoskeleton for long-running agents, with time-sliced execution, gated restarts, and versioned progress graphs. Evaluated on the 46-task LongHorizonBenchmark, with ~2.4× fewer input tokens and reward parity or better on 37 of 42 tasks.
An interactive survey of document understanding, spanning OCR, layout analysis, information extraction, and multimodal question answering.
An all-in-one RAG system that builds structured, graph-based memory using knowledge graphs.
I lead an open-source research community focused on Ancient Chinese and literature, with language models and datasets.
Sincerely looking for PhD positions for fall admission!
Table tennis. I enjoy the fast pace and tactical depth of the sport.
Martial arts. I am drawn to the discipline, tradition, and physical and mental cultivation.
GitHub
Google Scholar
LinkedIn
Zhihu
rednote
