I am a first-year PhD student at MAC Lab, Xiamen University, advised by Prof. Liujuan Cao and Assoc. Prof. Xiawu Zheng.
My research focuses on long-horizon visual understanding and multimodal models. I study how visual systems can select informative observations and maintain state over time, so they can understand objects, events, and their relationships in continuous video streams under limited compute and memory.
Building on this work, I want to explore the connection between visual understanding and world modeling: how a system can move from interpreting past events to modeling how its environment changes, and use prediction to support perception and reasoning. My long-term goal is visual intelligence that observes, updates, and reasons continuously, revising and enriching its understanding of the world as new evidence arrives.
Starting my first year as a PhD student!
Joined AMap, Alibaba Group, as a research intern.
🎉 1 paper (LLM Orchestration) has been accepted to ICML 2026.
🎉 1 paper (WFS-SB) has been accepted to CVPR 2026.
🎉 1 paper (QuoTA) has been accepted to AAAI 2026.
* Equal contribution
All 9 publications
2026CVPR 2026
Detects semantic boundaries from a noisy query-frame relevance signal, then allocates frame budgets by segment importance. The training-free method improves VideoMME, MLVU, and LongVideoBench by 5.5, 9.5, and 6.2 points.
2026AAAI 2026 · Equal contribution
Allocates visual tokens to frames before cross-modal interaction. CoT turns a complex query into scorable clues, improving six video benchmarks by 3.2 points on average under the same token budget.
2026arXiv preprint · Equal contribution
Partitions video into coherent events, localizes a query-relevant anchor in each event, and applies adaptive MMR for global refinement across coverage, relevance, and diversity.
Multimodal and video understanding
Began research in generative vision and facial aesthetics
A detailed account of Qwen3.5, from pixels and visual tokens to Gated DeltaNet, Full Attention, MRoPE, and MoE, with equations and tensor shapes for the 9B model.
75 min readHow Code-as-World connects visual evidence, executable world representations, and abductive discovery, and what its results do and do not establish about physical understanding in an open world.
25 min readA close reading of streaming video evaluation, incremental inference, persistent state, and thinking during observation, with a focus on evidence, compute budgets, and questions the system has not seen yet.
50 min readAn English reader's map to the full Chinese audit of video codecs, OneVision-Encoder, codec-video-prep, LLaVA-OneVision-2, and Mage-VL.
12 min read