晚期交互压缩的简单线性校正
原标题：Simple Linear Corrections for Compression in Late Interaction
AI 摘要
该论文研究晚期交互（多向量）检索模型在文档向量压缩后检索质量下降的问题，提出一种轻量级全局线性校正方法：冻结编码器与压缩器，通过蒸馏未压缩模型的 MaxSim 分数分布学习校正矩阵，可在搜索时或离线应用于存储向量。在 ColBERTv2、Jina v4（MS MARCO、BEIR）和 ColPali2（ViDoRe v3）上评估，域内设置下 160 个 BEIR 配置中有 154 个优于基线压缩，MS MARCO 迁移效果因模型和数据集而异，ViDoRe v3 上所有压缩器和预算均有提升。
正文节选
Simple Linear Corrections for Compression in Late Interaction Abstract Late-interaction (multi-vector) models represent documents using a set of token-level vectors, providing a middle ground between cross-encoders and dense encoders in terms of quality and cost of embedding. However, compared to the latter, the indices are costly to store and score. Pooling and pruning based compression techniques are deployed widely but deteriorate in quality noticeably with aggressive compression. We investig