A Deep Dive into Multi-Head Attention and Multi-Aspect Embedding

Abstract

<p>Multi-vector embedding models play an increasingly important role in retrievalaugmented generation, yet their internal behaviour lacks comprehensive analysis. We conduct a systematic, head-level study of the 32-head Semantic Feature Representation (SFR) encoder with the FineWeb corpus containing 10 billion tokens. For a set of 4,000 web documents, we pair head-specific embeddings with GPT-4o topic annotations and analyse the results using t-SNE visualisations, heat maps, and a 32-way logistic probe. The analysis shows that (i) clear semantic separation between heads emerges only at an intermediate layer, (ii) some heads align with specific topics while others capture broader corpus features, and (iii) naive pooling of head outputs can blur these distinctions, leading to frequent topic mismatches. The study offers practical guidance on where to extract embeddings, which heads may be pruned, and how to aggregate them to support more transparent and controllable retrieval pipelines.<br></p>