arXiv:2605.29384v1 Announce Type: cross 
Abstract: We propose Latent Terms, a method revealing that models trained for dense retrieval, whether single- or multi-vector, learn representations that can trivially be decomposed into retrieval-ready sparse features. When trained on frozen retrievers, Sparse Autoencoders without any retrieval-specific adjustments extract a latent vocabulary with approximately Zipfian collection statistics, directly suitable for classical sparse retrieval scoring via BM25. This approach enables sparse retrieval while requiring no learned expansion objective or sparse retrieval supervision whatsoever, and can be readily applied to any dense retriever. Latent Terms is able to match or outperform single-vector scoring methods from its own base model as well as comparable SPLADE variants. In addition, it substantially outperforms its base model on LIMIT, a task specifically designed to highlight the failures of single-vector retrieval. Overall, our results highlight that neural retrievers contain more expressive and indexable structure than their default scoring functions expose, but that other methods can nonetheless be leveraged.

تم اقتراح طريقة جديدة تُدعى المصطلحات الكامنة، تُظهر أن نماذج الاسترجاع الكثيف يمكن تفكيكها إلى ميزات نادرة جاهزة للاسترجاع، باستخدام مُشفّر نادر دون الحاجة إلى تعديلات تدريب إضافية. تُظهر هذه الطريقة أن هذه النماذج يمكن أن تستخرج بشكل فعال مفردات زيفية مناسبة لتقييم BM25.

Se ha propuesto un nuevo método llamado Términos Latentes, que demuestra que los modelos de recuperación densa pueden descomponerse en características dispersas listas para la recuperación, utilizando específicamente Autoencoders Dispersos sin requerir ajustes adicionales de entrenamiento. Este método muestra que estos modelos pueden extraer de manera efectiva un vocabulario Zipfiano adecuado para la puntuación BM25.

Une nouvelle méthode appelée Termes Latents a été proposée, démontrant que les modèles de récupération dense peuvent être décomposés en caractéristiques éparses prêtes pour la récupération, en utilisant spécifiquement des Autoencodeurs Épars sans nécessiter d'ajustements d'entraînement supplémentaires. Cette méthode montre que ces modèles peuvent efficacement extraire un vocabulaire Zipfien adapté au scoring BM25.

A new method called Latent Terms has been proposed, demonstrating that dense retrieval models can be decomposed into sparse features ready for retrieval, specifically utilizing Sparse Autoencoders without requiring additional training adjustments. This method shows that these models can effectively extract a Zipfian vocabulary suitable for BM25 scoring.