Skip to content
Hot eventLive

VLM mid-layer interpretability: encoded but disconnected errors

1 reports1 sources3 hr ago updated

Get the story

AI overview

2026-10-02 arXiv 论文对 LLaVA-1.5-7B、Qwen2.5-VL-7B、InternVL3-8B 三种架构在 POPE 基准上的失败模式做分解,发现 68–91% 的错误在中间层已编码正确答案,但残差流修补无法因果激活该信号;Qwen 每头 0/12,600 次翻转。

Generated from reports · updated 3 hr ago

Timeline

Follow the coverage from different angles.

Oct 2, 2026
  1. arXiv · Computer Vision
    Encoded but Disconnected: 视觉-语言模型失败模式分解研究

    在 LLaVA-1.5-7B、Qwen2.5-VL-7B、InternVL3-8B 三种架构上,POPE 基准中 68-91% 的错误在中间层编码了正确答案,但残差流修补无法因果激活该信号(Qwen 每头 0/12,600 次翻转)。

Heat trend

There is not enough continuous observation data to draw a trend yet.