Live content changes moment by moment. My team brings together video, audio, language, and creator-viewer interactions to build models for recommendation, content understanding, and feed-quality governance. I lead the technical direction and team development, connecting research priorities with deployment and measurable product impact. Our work spans four connected directions, supported by shared foundation models.
-
LLM-driven recommendation
Modeling evolving user interests with LLMs to guide retrieval and ranking, balancing familiar preferences with the discovery of new creators and live content.
-
Feed-quality governance
Combining multimodal issue detection, human feedback, and automated evaluation to improve content quality and safety decisions in live recommendation feeds.
-
Live content representation
Learning multimodal embeddings and semantic codes for live rooms, creators, and content slices, so retrieval and ranking can generalize across short-lived, continuously changing content.
-
Live understanding
Turning video, audio, language, and interactions into room and creator summaries, content categories, and contextual signals that make live experiences understandable to downstream models.
-
Live foundation models
Developing large and compact multimodal models for live content, with a focus on audio and vision-language understanding through continued pretraining and task adaptation. We are building shared domain knowledge from explicit evaluation rubrics and model reasoning, while optimizing inference over densely sampled video frames. Together, these efforts provide a foundation for recommendation, representation, live understanding, and feed-quality governance, balancing model capability with serving efficiency.
Experience & Impact
-
ByteDance · TikTok Live Content AI · Team Lead
Lead the team’s research and production efforts across multimodal foundation models, live content understanding, recommendation, and feed-quality governance, while setting technical direction and developing the team.
Core expertise: Research leadership · Multimodal foundation models · Large-scale AI systems
-
ByteDance · TikTok Live Content Understanding · Senior AI Researcher
Built multimodal understanding capabilities for live content and connected them with recommendation systems. The work covered real-time content representations, creator-viewer interaction, cold-start retrieval, content taxonomy, and quality governance.
Core expertise: Multimodal content understanding · Real-time recommendation · Representation learning · Interaction modeling
-
ByteDance · TikTok Video Content Understanding · AI Researcher
Developed multimodal pretraining and video understanding models for TikTok, bringing content representations into cold-start retrieval, recommendation diversification, and content governance while reducing reliance on manual labeling.
Core expertise: Multimodal pretraining · Cross-modal retrieval · Noisy-label learning · Recommendation embeddings
-
ByteDance · Xigua Video · AI Researcher
Developed video highlight detection for long-form content, using matched clips as supervision to scale training. Model-generated highlights were deployed in the recommendation feed, improving viewing time and active days in online experiments.
Core expertise: Video understanding · Highlight detection · Large-scale supervision mining · Online experimentation
-
Tencent Youtu Lab · Computer Vision Researcher
Worked on crowded pedestrian detection as part of a team that won the CrowdHuman challenge. Built fashion image retrieval from scratch and brought it to a public cloud service; developed smart-home video highlight detection deployed with two flagship customers. These projects also produced conference publications in detection, cross-modal retrieval, and video highlights. Built behavior recognition for content moderation.
Core expertise: Crowded object detection · Cross-modal retrieval · Video highlight detection · Action recognition · 0-to-1 product development
-
Megvii · Research Intern
Conducted research on object detection, with a focus on model development, experimentation, and evaluation.
Core expertise: Object detection · Model training · Experimental evaluation
Education
Fudan University
Computer Science and Technology · 2017–2020
Honor: Shanghai Outstanding Graduate · National Scholarship recipient
Research focus: Human-object interaction · Object detection · Few-shot learning · Zero-shot learning
Zhejiang University of Technology
Communication Engineering · 2013–2017
Honor: Zhejiang Provincial Outstanding Graduate · National Scholarship recipient
Projects: Designed and built a camera-guided autonomous racing car for the NXP Cup, earning first place in Zhejiang and a National First Prize · Initiated a separate project to design and build a miniature quadrotor from scratch
Focus: Intelligent control · Autonomous vehicles · Quadrotor systems