Kunxi Li

About Me

Hi! I am Kunxi Li, a graduated Master student from AI4GC Lab at Zhejiang University, advised by Prof. Shengyu Zhang.

My research interests include efficient multimodal large language models, long-context inference, KV Cache optimization, and knowledge migration.

I am currently pursuing a Ph.D. in Artificial Intelligence at the College of Computer Science and Technology, Zhejiang University.

During AI4GC

During my time at AI4GC Lab, I have been working on efficient inference and knowledge migration for large language models and multimodal large language models.

One line of my research focuses on KV Cache optimization for multimodal long-context inference. I studied the modality-specific characteristics of attention in multimodal large language models and designed adaptive KV Cache eviction strategies for efficient inference. This work was accepted by ACL 2025.

I also worked on audio-video large language model inference acceleration, where I explored adaptive focusing and cross-modal calibration for reducing redundant KV Cache while preserving important audio and visual information. This work led to AccKV, which was accepted by AAAI 2026.

Another line of my research focuses on knowledge migration across heterogeneous models, tasks, and modalities. I explored how knowledge can be transferred and integrated across different models and tasks through parameter-level migration and merging. This work led to MergeNet, which was accepted by AAAI 2025.

Now

Now: Ph.D. student @ Zhejiang University.