arXiv preprint / cs.CV

What Do Visual-Head Discovery Scores Actually Find? Attention Source, Necessity, and Random-Control Pitfalls in Vision-Language Models

作者：Md Abdul Kadir, Daniel Sonntag · 单位：German Research Center for Artificial Intelligence (DFKI), Oldenburg University · 会议/期刊：arXiv preprint · 方向：cs.CV · 发布日期：2026-10-12

2026-10-12视觉语言视觉感知规划控制
新标签打开 AlphaXiv
What Do Visual-Head Discovery Scores Actually Find? Attention Source, Necessity, and Random-Control Pitfalls in Vision-Language Models figure
AlphaXiv 中文概览（可滚动查看）