Publications

2026

1. KDD

   

   Towards Trustworthy Multimodal Moderation via Policy-Aligned Reasoning and Hierarchical Labeling

   Anqi Li, Wenwei Jin, Jintao Tong, Pengda Qin, Weijia Li, and Guo Lu^†

   In Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2026

   TL;DR: A policy-aligned reasoning framework with hierarchical labeling that makes multimodal content moderation transparent and trustworthy.

   First Author arXiv Code

   

   First-author publication

2. CVPR

   

   Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention

   Junhao Du, Jialong Xue, Anqi Li, Jincheng Dai, and Guo Lu^†

   In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026

   TL;DR: Unifies spatiotemporal pruning and clustering so Video-LLMs stay accurate while keeping only a few percent of visual tokens.

   arXiv

   
3. KDD

   

   UniNote: A Unified Embedding Model for Multimodal Representation and Ranking

   Jinghan Zhao, Wenwei Jin^†, Anqi Li, Jintao Tong, Luya Mo, Jiawei Li, Bin Li, and Yao Hu

   In ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2026

   TL;DR: A unified multimodal embedding model with contrastive SFT and RL-based ranking refinement, deployed for industrial item-to-item retrieval at Xiaohongshu.

   arXiv
4. PR

   

   You Can Mask More For Extremely Low-Bitrate Image Compression

   Anqi Li, Feng Li, Jiaxin Han, Huihui Bai, Runmin Cong, Chunjie Zhang, Meng Wang, Weisi Lin, and Yao Zhao^†

   Pattern Recognition, 2026

   TL;DR: Masks more of the image than prior methods and lets a learned model inpaint it back, pushing image compression to extremely low bitrates.

   First Author arXiv Code

   

   First-author publication

2025

1. ICLR

   

   Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation

   Anqi Li, Feng Li, Yuxi Liu, Runmin Cong, Yao Zhao, and Huihui Bai^†

   In International Conference on Learning Representations, 2025

   TL;DR: One generative codec that adapts compression granularity on the fly, covering a wide range of bitrates with a single model.

   First Author arXiv Code

   

   First-author publication

2. NeurIPS

   

   FlowCut: Rethinking Redundancy via Information Flow for Efficient Vision-Language Models

   Jintao Tong, Wenwei Jin, Pengda Qin, Anqi Li, Yixiong Zou, Yuhong Li, Yuhua Li, and Ruixuan Li^†

   In Advances in Neural Information Processing Systems, 2025

   TL;DR: Rethinks visual token redundancy through information flow and cuts tokens where information stops flowing, accelerating VLMs with minimal accuracy loss.

   arXiv Code