Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval

Seongwan Park, Taeklim Kim, Youngjoong KoEMNLP 2025

TL;DR

This paper introduces a novel framework using Sparse Autoencoders (SAEs) to interpret dense retrieval models by decomposing their opaque embeddings into distinct, human-understandable latent concepts. This allows for better understanding of how these models work and leads to a new retrieval paradigm, Concept-Level Sparse Retrieval (CL-SR), which offers comparable performance to dense models with significantly improved efficiency and robustness to vocabulary/semantic mismatches.

What Problem Does This Paper Address?

Traditional information retrieval (IR) methods, like keyword matching, suffer from vocabulary and semantic mismatches. Modern dense retrieval (DR) models, powered by deep learning, overcome these issues by representing queries and documents as dense vectors in a continuous space, leading to state-of-the-art performance. However, these dense embeddings are notoriously difficult to interpret, making it hard to understand why a particular document is retrieved or how the model makes its similarity judgments. This lack of transparency hinders trust and debugging, especially in critical applications. The paper aims to bridge this gap by making dense retrieval models more interpretable and leveraging this interpretability for a more efficient and robust retrieval system.

Methods

The core of the proposed method involves using Sparse Autoencoders (SAEs). An SAE is trained to reconstruct dense embeddings from a pre-trained Dense Passage Retrieval (DPR) model. The SAE's encoder maps the dense embedding to a sparse latent representation, where only a subset of "latent concepts" are activated. The decoder then attempts to reconstruct the original dense embedding from these sparse latents.

Key aspects of their method include:

  • BatchTopK SAE: A specific SAE variant is used, which allows for flexible activation of latents within a batch.
  • Latent Concept Interpretation: Natural language descriptions are generated for each activated latent concept using a powerful LLM (GPT-4.1 mini), allowing humans to understand what each concept represents.
  • Concept-Level Sparse Retrieval (CL-SR): This is a novel retrieval framework that uses the extracted latent concepts as the fundamental units for indexing and retrieval. Instead of traditional term-based sparse retrieval or opaque dense retrieval, CL-SR builds an inverted index of these semantically generalized latent concepts.
  • IDF Weighting: To mitigate the influence of high-frequency, abstract latents, Inverse Document Frequency (IDF) is applied to the latent activation values, boosting the importance of more specific concepts.

Datasets

  • MSMARCO: Used for training SAEs and evaluating retrieval performance. Approximately 8.8 million passages and 0.5 million training queries were used.
  • TREC DL 2019 & 2020: Used for evaluating retrieval performance.
  • TREC-COVID: Used for evaluating the generalizability of SAEs to unseen datasets.
  • NFCorpus: Also used for evaluating SAE generalizability to unseen datasets.

The paper does not introduce new benchmarks but rather evaluates their methods on existing, widely-used IR datasets.

Key Insights and Ideas

  • Dense Embeddings are Decomposable: The paper demonstrates that dense embeddings, despite their complexity, can be effectively decomposed into a set of distinct, interpretable latent concepts using SAEs.
  • Latent Concepts as Semantic Units: These extracted latent concepts act as semantically generalized units, capable of capturing meaning beyond exact lexical matches.
  • Interpretability Enables Better Retrieval: By understanding the latent concepts, the authors developed CL-SR, a retrieval system that leverages these concepts. This approach bridges the gap between the semantic richness of dense embeddings and the efficiency of sparse retrieval.
  • Robustness to Mismatches: CL-SR's reliance on semantically generalized concepts makes it inherently more robust to vocabulary and semantic mismatches compared to traditional sparse methods.
  • Efficiency Gains: Representing documents with a compact set of latent concepts significantly reduces computational costs (FLOPs) and index storage size.

Key Results

  • Retrieval Performance: CL-SR (Max configuration) achieved competitive retrieval accuracy, on par with or exceeding other neural sparse retrieval baselines on MSMARCO Dev (MRR@10 of 0.368, Recall@1k of 0.969). On TREC-DL datasets, it also showed strong performance, despite these datasets often favoring long-tailed queries.
  • Efficiency: CL-SR (Efficient configuration) achieved comparable performance to SPLADE v2 with significantly lower FLOPs per query (0.11) and index storage (0.57 GB).
  • Interpretability: Human annotators achieved high accuracy (0.943) in identifying passages based on latent descriptions, confirming the effectiveness of the generated explanations. They also showed high accuracy (over 0.9) in simulating the DPR model's ranking behavior using latent descriptions.
  • Robustness: On a challenging "Mismatch Set" (queries where BM25 fails), CL-SR demonstrated significantly smaller performance drops compared to traditional sparse models, highlighting its robustness to vocabulary and semantic mismatches.

Limitations and Future Work

  • SAE Limitations: The authors acknowledge known limitations of SAEs, such as features being incomplete, dataset-dependent, and potentially polysemantic (one latent concept representing multiple meanings). They stick to a widely used SAE variant and suggest exploring newer architectures in future work.
  • LLM Evaluation Sensitivity: Evaluating the interpretability of SAEs primarily relies on LLMs, which can be sensitive to prompts and have variable reasoning. While they compensated with human studies, this method is labor-intensive and difficult to scale.
  • Future Work: The authors suggest exploring newer SAE variants, applying their framework to other DPR models, and further investigating the impact of latent concept counts on CL-SR performance. They also believe their framework can be used to improve dense retrieval models or enhance neural sparse retrievals.

Summary updated

BibTeX
@inproceedings{park2025decoding,
  author = {Seongwan Park and Taeklim Kim and Youngjoong Ko},
  title = {{Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure