arXiv：cs.CL· Korbinian Randl, Guido Rocchietti, Aron Henriksson, Ziawasch Abedjan, Tony Lindgren, John Pavlopoulos·· 3 小时前
RAG-E：用局部解释量化 RAG 中检索器与生成器的对齐程度
Quantifying Retriever-Generator Alignment in RAG with Local Explanations
AI 导读
研究者提出端到端可解释性框架 RAG-E，通过归因方法量化 RAG 中检索器与生成器的对齐程度，并提出 WARG 指标衡量生成器对文档的使用与检索器排序的一致程度。在 PopQA、QAMPARI 和 TREC CAST 数据集上的实验显示，生成器常忽略排名靠前的文档，转而依赖排名较低的文档。WARG 比 Pearson 和 Spearman 相关性更能反映这种对齐，并可指示 RAG 性能。
正文
Abstract:Retrieval-Augmented Generation (RAG) systems combine dense retrievers and language models to ground outputs in external documents. However, the interaction between these components remains opaque, creating challenges for deployment in high-stakes domains. We present RAG-E, an end-to-end explainability framework that quantifies retriever-generator alignment through mathematically grounded attribution methods. Our approach adapts Integrated Gradients for retriever analysis, proposes a Monte Carlo-stabilized Shapley Value approximation for generator attribution, and introduces the Weighted Alignment between Retriever and Generator (WARG) metric to measure how closely the generator's document usage aligns with retriever rankings. Experiments on PopQA, QAMPARI, and TREC CAST datasets reveal substantial misalignment: depending on the model and setting, generators often ignore top-ranked documents and rely on documents ranked as less relevant. We show that WARG captures retriever-generator alignment better than Pearson and Spearman correlations and can serve as an indicator of RAG performance. RAG-E and WARG provide a practical framework for auditing this interaction, enabling more reliable and transparent RAG systems.
| Comments: | Accepted at EMNLP 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2601.21803 [cs.CL] |
| (or arXiv:2601.21803v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2601.21803 arXiv-issued DOI via DataCite |
Submission history
From: Korbinian Randl [view email]
[v1]
Thu, 29 Jan 2026 14:47:00 UTC (1,954 KB)
[v2]
Tue, 7 Jul 2026 09:51:50 UTC (979 KB)
[v3]
Wed, 7 Oct 2026 19:36:30 UTC (1,145 KB)
来源：arXiv：cs.CL · arxiv.org