Xetrieval: Explaining Dense Retrieval through Feature Elicitation

Feature Elicitors and a shared sparse dictionary produce retrieval explanations evaluated through pair-level and task-level interventions — **Overview.** Feature Elicitors make document semantics more explicit through auxiliary views, and a shared SAE identifies describable features shared by queries and documents. Pair-level interventions assess their influence on relevance scores. Task-level feature modulation evaluates retrieval steering with SAE-decoded representations.

Abstract

Explaining dense retrieval through feature elicitation

Dense retrieval enables effective semantic matching, yet the semantic factors underlying its relevance scores remain opaque. Sparse autoencoders (SAEs) support feature-level explanations by decomposing dense embeddings into sparse semantic features. However, the document semantics underlying a retriever's relevance scores may not be sufficiently explicit in the embeddings to be recovered as interpretable SAE features.

We introduce Xetrieval, a framework for explaining dense retrieval through feature elicitation. Feature Elicitors make retrieval-relevant aspects of document meaning more explicit through auxiliary views, facilitating their identification as interpretable SAE features. A shared sparse dictionary learns from these varied representations to express query–document relations through describable features. We assess their influence on relevance scores through interventions on the original document embeddings.

Experiments show that Xetrieval yields more interpretable features than an SAE baseline and identifies feature sets with a stronger influence on the retriever's scores. Task-level feature modulation further demonstrates more effective retrieval steering. These findings highlight the value of feature elicitation for relevance attribution and retrieval steering.

Feature elicitation

Auxiliary document views make retrieval-relevant aspects of document meaning more explicit for sparse decomposition.

Interpretable sparse features

A shared sparse dictionary learns from varied semantic expressions and supplies describable features for query–document relations.

Attribution and steering

Pair-level interventions assess influence on the original retrieval scores; task-level modulation evaluates retrieval steering.

Experiments

Feature interpretability, attribution, and retrieval steering

We first assess sparse feature quality and interpretability, then examine the selected features' influence on relevance scores and their utility for task-level steering. These analyses jointly assess the quality of retrieval explanations. We also measure the computational overhead of feature elicitation.

Sparse Dictionary Quality

Balancing semantic consistency and reconstruction quality.

We compare seven SAE variants using reconstruction error, intruder detection accuracy, and retrieval with reconstructed embeddings. TopK at k = 256 offers a favorable balance across these criteria and is used in the subsequent explanation experiments.

Comparison of seven SAE variants across sparsity levels — Reconstruction error, intruder detection accuracy, and retrieval with reconstructed embeddings. Reconstruction and retrieval are evaluated on BRIGHT Biology.

Structure of Sparse Representations

Feature elicitation broadens the features available for explanation.

Xetrieval achieves lower reconstruction error and activates more distinct dictionary features than the SAE baseline. This broadens the set of candidate features available for retrieval explanation. Active Feature Count measures the number of distinct features activated at least once across the evaluated embedding set.

Reconstruction error and distinct active feature counts for SAE and Xetrieval — Reconstruction error and distinct active feature counts on BRIGHT for SAE and Xetrieval.

Interpretable Sparse Features

Feature descriptions better match observed activations.

In the automated Detection evaluation, a language model uses each feature description to distinguish activating from non-activating texts. Xetrieval's scores shift higher relative to SAE and Untrained SAE, indicating closer agreement between its descriptions and the evaluated activation patterns.

Detection score distributions — Detection Score distributions for Xetrieval, SAE, and Untrained SAE.

Pair-Level Interventions

Selected feature spans have greater influence on relevance scores.

With matched feature counts and a fixed query embedding, erasing Xetrieval's selected decoder spans from the original document embedding causes larger average similarity decreases than SAE. Retaining these spans yields larger increases. These responses provide behavioral support for relevance attribution through the selected feature sets.

Pair-level document-side intervention results — Erasure and retention of selected feature spans in original document embeddings. The non-overlapping control uses Xetrieval's dictionary.

Task-Level Interventions

Task-level feature selection supports retrieval steering.

We select features associated with task relevance and scale their activations before decoding. On most datasets, suppressing Xetrieval's key features reduces retrieval performance, while amplifying them maintains or improves it. Relative to SAE, Xetrieval generally yields larger losses under suppression and greater gains under amplification.

Task-level feature modulation for SAE and Xetrieval on seven datasets — Retrieval with SAE-decoded representations after suppressing or amplifying Key and Other feature sets. The reference α = 1 uses unmodified sparse activations.

Computational Efficiency

Feature elicitation adds little processing latency.

End-to-end timing on the BRIGHT Biology document set shows a small absolute latency overhead for Xetrieval. Each Feature Elicitor requires one MLP forward pass, and the shared SAE can encode the additional views in a batch. Both operations act on fixed-dimensional embeddings.

Processing latency of the baseline and Xetrieval as document count increases — Processing latency on BRIGHT Biology, including feature elicitation and SAE encoding of the auxiliary views.

Code

Open-source inference code

The release contains the inference pipeline, shared TopK SAE checkpoint, and Feature Elicitor checkpoints for qa, summary, and purpose document-side views.

git clone https://github.com/Hihiczx/Xetrieval.git xetrieval
cd xetrieval
pip install -r requirements.txt

python explain_retrieval.py \
  --input_jsonl examples/query_doc_pairs.jsonl \
  --output_jsonl outputs/explanations.jsonl \
  --embedding_model_name intfloat/e5-large-v2 \
  --feature_elicitor_dir checkpoints/feature_elicitors \
  --sae_checkpoint checkpoints/sae_model.pt \
  --device cuda \
  --batch_size 64

Citation

Cite Xetrieval

If you find Xetrieval useful for your research, please cite:

@article{cai2026xetrieval,
  title={Xetrieval: Mechanistically Explaining Dense Retrieval},
  author={Cai, Zhixin and Bai, Jun and Liu, Yang and Li, Jiaqi and Zhang, Yichi and Li, Taichuan and Chen, Zhuofan and Jia, Zixia and Zheng, Zilong and Rong, Wenge},
  journal={arXiv preprint arXiv:2605.29507},
  year={2026}
}