Zhiyuan Yan – PhD Student at Peking University

Zhiyuan Yan

📖 PhD Student, 2024.09-2028.07(expected)

🏫 Peking University, previously at CUHK-SZ

💡 Multimodal, AIGC Detection, AIGC, AI4Science

About Me

I am a second-year CS PhD Student at Peking University (PKU), supervised by Prof. Li Yuan (Deep Learning) and Prof. Fanyang Mo (AI4Science). I have published 10+ CCF-A papers (as the first/co-first author) at the top international AI conferences with a total Google Scholar citations of 1000+, including an oral paper at ICML 2025 (top 1%).

I have extensive industrial experience in multimodal foundation models. I am currently part of the Meituan LongCat Multimodal Foundation Model Team through the Beidou Program, where I am responsible for the development of the native unified multimodal system. Previously, I worked at Baidu ERNIE (worked directly with the CTO) through the “ERNIE Star Program,” and before that, I was a research intern at Tencent Youtu Lab and Tencent AI Lab.

My current research interests are: (1) Multimodal, especially unified multimodal generation and understanding: My representative work in this area is UAE, a unified multimodal system with large-scale long-context SFT and RL-based post-training designed for unified multimodal generation and understanding. Through this journey, I gain extensive hands-on experience, deep insights and understanding into SFT, RL, and data curation for unified models. (2) AI for Science: Leveraging the advanced multimodal techniques to address the key challenges in the science domains, especially organic chemistry like structure elucidation by spectroscopy and molecular representation modeling.

My previous research direction is AIGC Detection and Deepfake detection: I have developed generalizable and interpretable methods for detecting AI-generated images and videos. My work includes six first-author papers accepted at CCF-A venues. My impactful contribution, DeepfakeBench, integrates 37 SOTA detection methods and 13 datasets into a unified modular codebase with standardized training and evaluation protocols, with near 1k stars on GitHub.