Xetrieval: Mechanistically Explaining Dense Retrieval

Researchers introduce Xetrieval, a mechanistic framework that explains how dense retrieval models assign relevance scores by decomposing high-dimensional embeddings into interpretable features. The method uses a lightweight reasoning internalizer to enrich embeddings with reasoning information and provides human-readable feature-level explanations of retrieval decisions, advancing transparency in neural information retrieval systems.

Xetrieval addresses a fundamental challenge in modern machine learning: the interpretability of neural retrieval systems. Dense retrievers, which power search engines and recommendation systems, make relevance decisions through opaque high-dimensional embeddings that resist human interpretation. This paper proposes a solution by introducing a mechanism that bridges the gap between learned representations and human-understandable explanations, representing meaningful progress in explainable AI for information retrieval.

The framework operates at the embedding level rather than relying on surface-level signals like keyword matches or post-hoc text generation. By implementing a lightweight reasoning internalizer that performs Chain-of-Thought approximations within embedding space, Xetrieval avoids computationally expensive autoregressive generation while maintaining rich semantic information. The method then decomposes these enriched embeddings into sparse, interpretable features with natural language descriptions—a crucial step for understanding what information drives retrieval decisions.

For the broader AI community, this work has significant implications. As dense retrievers become increasingly central to applications from search to question-answering systems, the ability to explain their decisions becomes critical for debugging, auditing, and improving these systems. The research demonstrates that feature-level explanations enable stronger intervention effects and task-level feature steering, suggesting practical applications beyond pure interpretability.

The availability of source code democratizes access to mechanistic interpretability techniques, potentially inspiring similar approaches in other neural architectures. Future work might extend these methods to other embedding-based systems or combine them with other explainability techniques for even richer understanding of neural behavior.

• →Xetrieval decomposes dense retrieval embeddings into human-interpretable features with natural language descriptions
• →A lightweight reasoning internalizer enriches embeddings with Chain-of-Thought information without expensive text generation
• →Feature-level explanations enable intervention effects and task-level steering of retrieval behavior
• →The framework operates at the embedding level, providing deeper insights than surface-level lexical matching explanations
• →Open-source release enables broader adoption of mechanistic interpretability techniques for neural retrieval systems