Xetrieval: Mechanistically Explaining Dense Retrieval

Xetrieval is a mechanistic explanation framework for dense retrieval that adds reasoning signals to sentence embeddings.

It decomposes embeddings into sparse, human-readable features and aggregates them to explain retrieval scores.

Experiments show the extracted features are coherent, interventions are stronger, and retrieval can be steered at the task level.

The approach aims to make dense retrievers more interpretable for debugging, analysis, and controllable retrieval.