RAG Failures Explained Through Circuit-Level Attribution Graphs

A new study from arXiv (2605.14192) investigates why Retrieval-Augmented Generation (RAG) systems produce incorrect answers despite access to external evidence. Using circuit tracing, researchers constructed attribution graphs that model information flow through transformer layers during decoding. These graphs reveal interactions among retrieved context, intermediate activations, and generated tokens. Across multiple question-answering benchmarks, consistent structural differences emerged: correct predictions exhibit deeper reasoning pathways than incorrect ones. The work provides a model-internal, circuit-level view of how external evidence is integrated into reasoning, offering a novel perspective on RAG failures.