Persistent Object Narratives for Token-Efficient Video Language Models
arXiv preprint, 2026
Compact identity and state tokens connect recurring objects across time, providing a token-efficient interface for Video-LLMs.
Overview & resourcesAbout Me
I am a Ph.D. student in Computer Science at Tianjin University, advised by Xiaojie Guo (郭晓杰). Previously, I received my M.S. in Computer Science from Johns Hopkins University (January 2024–May 2025) and my B.Eng. in Software Engineering from Beijing Jiaotong University (2019–2023).
My research focuses on computer vision, video understanding, and vision-language models. I am particularly interested in how models preserve object identity and state over time, reason from visual evidence, and represent videos efficiently. My earlier work includes tiny-object detection and mixed-precision quantization.
I welcome discussions about research and potential collaborations. Please feel free to contact me.
News
- SlotNarrative, our work on token-efficient video-language models, is available on arXiv.
- We introduced TOC-Bench, a benchmark for temporal object consistency in Video-LLMs.
Research Highlights
All publications
TOC-Bench: A Temporal Object Consistency Benchmark for Video Large Language Models
arXiv preprint, 2026
A diagnostic benchmark for object identity, state, and temporal continuity, grounded in object tracks and ordered visual evidence.
Overview & resourcesEducation
- Tianjin UniversityCurrent
Ph.D. student in Computer Science
- Johns Hopkins University–
M.S. in Computer Science
- Beijing Jiaotong University2019–2023
B.Eng. in Software Engineering
Industry Experience
- Huizhi Square Technology — Algorithm Engineer
- OPPO — Deep Learning Intern
- Sense Technology — Computer Vision Algorithm Intern
Honors & Awards
- Third Prize, Beijing Big Data Skills Competition, 2023
- Outstanding Undergraduate Thesis, Beijing Jiaotong University, 2023
- 4th Place, 4th International Competition on Human Identification at a Distance, 2023