LLM2IR: simple unsupervised contrastive learning makes long-context LLM great retriever

TL;DR

This paper presents LLM2IR, an unsupervised contrastive learning method that efficiently transforms decoder-only large language models into effective information retrieval systems, demonstrating strong performance across multiple benchmarks.

Contribution

Introduces a simple unsupervised contrastive learning framework to convert any decoder-only LLM into an IR model, revealing the link between context length and IR capability.

Findings

Models with longer context lengths have better IR performance.

LLM2IR achieves competitive results on multiple IR benchmarks.

Unsupervised training simplifies IR model development.

Abstract

Modern dense information retrieval (IR) models usually rely on costly large-scale pretraining. In this paper, we introduce LLM2IR, an efficient unsupervised contrastive learning framework to convert any decoder-only large language model (LLM) to an information retrieval model. Despite its simplicity, the effectiveness is proven among different LLMs on multiple IR benchmarks including LoCo, LongEmbed and BEIR. We also find that models with a longer context length tend to have a stronger IR capacity by comparing task performances of models in the same model family. Our work not only provides an effective way to build IR models on the state-of-the-art LLMs, but also shed light on the relationship between information retrieval ability and model context length, which helps the design of better information retrievers.