SAGE-CP：自适应VLM集成中边际覆盖掩盖路由条件失败

AI 摘要

论文提出 SAGE-CP 方法，针对自适应视觉语言模型（VLM）集成中边际覆盖可能掩盖特定路由条件失败的问题。作者证明，在单一 conformal 阈值下，某条路由的覆盖率可能降至零，而整体预测器仍保持边际有效。SAGE-CP 通过在开发数据上学习并冻结路由策略，再对早退和升级输入分别进行 Mondrian 分组校准，在十个基准上把平均最差路由覆盖率从较低水平提升，并缩小路由间覆盖差距，同时保持比全集成更小的预测集和更低的延迟与能耗。

正文节选

SAGE-CP: When Marginal Coverage Hides Route-Conditional Failures in Adaptive VLM Ensembles Abstract Adaptive vision-language model (VLM) ensembles save computation by stopping early on easy images and escalating harder ones. Although a single conformal threshold can make the final predictor marginally valid, it can still fail severely along one of its computational routes. We characterize this failure precisely: if a route has prevalence , marginal coverage allows its conditional coverage to fal

发布时间：2026-10-21 12:00

抓取时间：2026-10-20 12:18

来源机构：arXiv