CV
Curriculum vitae.
Contact Information
| Name | Anqi Li |
| Professional Title | Ph.D. Student |
| anqi.li@sjtu.edu.cn | |
| Phone | +86 15221070366 |
Professional Summary
Ph.D. student at Shanghai Jiao Tong University. Research interests include VLMs, agents, trustworthy reasoning, and efficient visual representations.
Experience
-
2025 - present Shanghai, China
Research Algorithm Intern, Frontier Exploration Center
Shanghai AI Laboratory
- Work on entity consistency in multi-shot video generation, including difficult evaluation scenarios with shot changes, viewpoint shifts, and multi-person interactions.
- Explore entity-centric hierarchical memory mechanisms for injecting explicit entity states into generative processes.
-
2024 - 2025 Beijing, China
Research Algorithm Intern, Content Understanding Group
Xiaohongshu
- Modeled business moderation policies as a hierarchical decision space for multi-level content safety labels.
- Designed tree-structured label representations and reinforcement learning objectives for path-level decision consistency.
- Related work accepted to KDD 2026.
Education
-
2025 - 2029 Shanghai, China
Ph.D. Student
Shanghai Jiao Tong University
Information and Communication Engineering
- {“Advisors”=>”Prof. Guo Lu and Prof. Wenjun Zhang.”}
-
2022 - 2025 Beijing, China
M.S.
Beijing Jiaotong University
Information and Communication Engineering
- {“Advisors”=>”Prof. Huihui Bai and Prof. Yao Zhao.”}
-
2018 - 2022 Wuhan, China
B.S.
Wuhan University of Technology
Software Engineering
Publications
-
2026 Towards Trustworthy Multimodal Moderation via Policy-Aligned Reasoning
KDD
A policy-aligned reasoning framework for fine-grained multimodal moderation with hierarchical labels and path-level reinforcement learning.
-
2025 Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation
ICLR
A controllable generative image compression framework that supports arbitrary compression rates with a single trained model.
-
2025 Flowcut: Rethinking Redundancy via Information Flow for Efficient Vision-Language Models
NeurIPS
An information-flow-based token pruning method for efficient vision-language model inference.
-
2026 Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention
CVPR
A unified spatial-temporal token compression framework for efficient long-video reasoning with Video-LLMs.
Awards
-
2023 First-Class Scholarship
Beijing Jiaotong University
Top 20%.
-
2023 Bronze Award, World Intelligent Driving Challenge
WIDC
Autonomous driving simulation track.
-
2025 Outstanding Graduate and Outstanding Master's Thesis
Beijing Jiaotong University
Top 10%.