|国家预印本平台
首页|Towards Understanding How Knowledge Evolves in Large Vision-Language Models

Towards Understanding How Knowledge Evolves in Large Vision-Language Models

Towards Understanding How Knowledge Evolves in Large Vision-Language Models

来源:Arxiv_logoArxiv
英文摘要

Large Vision-Language Models (LVLMs) are gradually becoming the foundation for many artificial intelligence applications. However, understanding their internal working mechanisms has continued to puzzle researchers, which in turn limits the further enhancement of their capabilities. In this paper, we seek to investigate how multimodal knowledge evolves and eventually induces natural languages in LVLMs. We design a series of novel strategies for analyzing internal knowledge within LVLMs, and delve into the evolution of multimodal knowledge from three levels, including single token probabilities, token probability distributions, and feature encodings. In this process, we identify two key nodes in knowledge evolution: the critical layers and the mutation layers, dividing the evolution process into three stages: rapid evolution, stabilization, and mutation. Our research is the first to reveal the trajectory of knowledge evolution in LVLMs, providing a fresh perspective for understanding their underlying mechanisms. Our codes are available at https://github.com/XIAO4579/Vlm-interpretability.

Sudong Wang、Yunjian Zhang、Yao Zhu、Jianing Li、Zizhe Wang、Yanwei Liu、Xiangyang Ji

计算技术、计算机技术语言学

Sudong Wang,Yunjian Zhang,Yao Zhu,Jianing Li,Zizhe Wang,Yanwei Liu,Xiangyang Ji.Towards Understanding How Knowledge Evolves in Large Vision-Language Models[EB/OL].(2025-03-31)[2025-05-12].https://arxiv.org/abs/2504.02862.点此复制

评论