智能体VLM流水线中的视觉编排税：审计与认证视觉证据复用
原标题：Visual Orchestration Tax in Agentic VLM Pipelines: Auditing and Certifying Visual Evidence Reuse
AI 摘要
该论文提出「视觉编排税」概念，指出在智能体式 VLM 流水线中，同一静态图像证据会被多个专家智能体反复重建为图像条件请求，造成编排层冗余。作者开发了从测量到认证的框架：审计侧用指标统计视觉证据触碰次数与结构冗余，在 SeeingEye 和 MAMMQA 等任务上发现 66.8%–75.6% 的触碰冗余；认证侧提出 SharedVisCache，按图像内容、预处理指纹和编码器假设进行契约感知的证据复用。实验显示契约验证可将 75.0%–75.5% 的重复触碰认证为可复用，同时保持输出字符串不变，并在物理层将命中次数从 800 降至 200。
正文节选
Visual Orchestration Tax in Agentic VLM Pipelines: Auditing and Certifying Visual Evidence Reuse Abstract Agentic VLM pipelines increasingly pass the same static visual evidence through multiple specialist agents and tools. This design creates an orchestration-level redundancy mode: semantically unchanged images are repeatedly reconstructed as image-conditioned requests at the VLM API boundary. We call this phenomenon visual orchestration tax and develop a measurement-to-certification framework