State Space Models are Strong Text Rerankers

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Xu, Zhichao, Yan, Jinghua, Gupta, Ashim, Srikumar, Vivek
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910915978330112
author Xu, Zhichao
Yan, Jinghua
Gupta, Ashim
Srikumar, Vivek
author_facet Xu, Zhichao
Yan, Jinghua
Gupta, Ashim
Srikumar, Vivek
contents Transformers dominate NLP and IR; but their inference inefficiencies and challenges in extrapolating to longer contexts have sparked interest in alternative model architectures. Among these, state space models (SSMs) like Mamba offer promising advantages, particularly $O(1)$ time complexity in inference. Despite their potential, SSMs' effectiveness at text reranking -- a task requiring fine-grained query-document interaction and long-context understanding -- remains underexplored. This study benchmarks SSM-based architectures (specifically, Mamba-1 and Mamba-2) against transformer-based models across various scales, architectures, and pre-training objectives, focusing on performance and efficiency in text reranking tasks. We find that (1) Mamba architectures achieve competitive text ranking performance, comparable to transformer-based models of similar size; (2) they are less efficient in training and inference compared to transformers with flash attention; and (3) Mamba-2 outperforms Mamba-1 in both performance and efficiency. These results underscore the potential of state space models as a transformer alternative and highlight areas for improvement in future IR applications.
format Preprint
id arxiv_https___arxiv_org_abs_2412_14354
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle State Space Models are Strong Text Rerankers
Xu, Zhichao
Yan, Jinghua
Gupta, Ashim
Srikumar, Vivek
Computation and Language
Information Retrieval
Transformers dominate NLP and IR; but their inference inefficiencies and challenges in extrapolating to longer contexts have sparked interest in alternative model architectures. Among these, state space models (SSMs) like Mamba offer promising advantages, particularly $O(1)$ time complexity in inference. Despite their potential, SSMs' effectiveness at text reranking -- a task requiring fine-grained query-document interaction and long-context understanding -- remains underexplored. This study benchmarks SSM-based architectures (specifically, Mamba-1 and Mamba-2) against transformer-based models across various scales, architectures, and pre-training objectives, focusing on performance and efficiency in text reranking tasks. We find that (1) Mamba architectures achieve competitive text ranking performance, comparable to transformer-based models of similar size; (2) they are less efficient in training and inference compared to transformers with flash attention; and (3) Mamba-2 outperforms Mamba-1 in both performance and efficiency. These results underscore the potential of state space models as a transformer alternative and highlight areas for improvement in future IR applications.
title State Space Models are Strong Text Rerankers
topic Computation and Language
Information Retrieval
url https://arxiv.org/abs/2412.14354