BTR: Binary Token Representations for Efficient Retrieval Augmented Language Models

TL;DR

This paper introduces Binary Token Representations (BTR), a method using 1-bit vectors to significantly speed up retrieval-augmented language models and reduce storage needs while maintaining high accuracy.

Contribution

BTR is a novel approach that precomputes token representations with 1-bit vectors, enabling faster inference and lower storage for large language models.

Findings

BTR accelerates inference by up to 4x on five NLP tasks.

BTR reduces storage requirements by over 100x.

BTR maintains over 95% task performance.

Abstract

Retrieval augmentation addresses many critical problems in large language models such as hallucination, staleness, and privacy leaks. However, running retrieval-augmented language models (LMs) is slow and difficult to scale due to processing large amounts of retrieved text. We introduce binary token representations (BTR), which use 1-bit vectors to precompute every token in passages, significantly reducing computation during inference. Despite the potential loss of accuracy, our new calibration techniques and training objectives restore performance. Combined with offline and runtime compression, this only requires 127GB of disk space for encoding 3 billion tokens in Wikipedia. Our experiments show that on five knowledge-intensive NLP tasks, BTR accelerates state-of-the-art inference by up to 4x and reduces storage by over 100x while maintaining over 95% task performance.