LLM2Vec abstract

[Bridge] [Verse 1] Large decoder-only language models LLMs are the state-of-the-art models on most of today’s NLP tasks [Verse 2] Yet, the community is only slowly adopting these models for text embedding tasks which require rich contextualized representations. [Verse 3] In this work, we introduce LLM2Vec, a simple unsupervised approach that can transform any decoder-only LLM into a strong text encoder. [Chorus] LLM2Vec consists of three simple steps: One. Enabling bidirectional attention Two. Masked next token prediction, and Three. unsupervised contrastive learning. [Verse 4] We demonstrate the effectiveness of LLM2Vec by applying it to three popular LLMs [Verse 5] We outperform encoder-only models by a large margin on word-level tasks and reach a new unsupervised state-of-the-art performance on the Massive Text Embeddings Benchmark MTEB. [Verse 6] Moreover, when combining LLM2Vec with supervised contrastive learning, we achieve state-of-the-art performance on MTEB [Chorus] LLM2Vec consists of three simple steps: One. Enabling bidirectional attention Two. Masked next token prediction, and Three. unsupervised contrastive learning. Three. unsupervised contrastive learning. []Verse 7] Our strong empirical results and extensive analysis demonstrate that LLMs can be effectively transformed [Verse 8] into universal text encoders in a parameter-efficient manner without the need for expensive adaptation or synthetic GPT-4 generated data. [Chorus] LLM2Vec consists of three simple steps: One. Enabling bidirectional attention Two. Masked next token prediction, and Three. unsupervised contrastive learning.