Atlas: Few-shot Learning with Retrieval Augmented Language Models

TL;DR

Atlas demonstrates that retrieval-augmented language models can excel at knowledge-intensive tasks with minimal training data, outperforming larger models in few-shot settings by effectively leveraging document retrieval.

Contribution

The paper introduces Atlas, a pre-trained retrieval-augmented language model that achieves strong few-shot performance on knowledge-intensive tasks, with efficient document index updates.

Findings

Atlas reaches over 42% accuracy on Natural Questions with only 64 examples.

It outperforms a 540B parameter model by 3% in few-shot Natural Questions.

Retrieval-augmented models can effectively learn with minimal data and adaptable document content.

Abstract

Large language models have shown impressive few-shot results on a wide range of tasks. However, when knowledge is key for such results, as is the case for tasks such as question answering and fact checking, massive parameter counts to store knowledge seem to be needed. Retrieval augmented models are known to excel at knowledge intensive tasks without the need for as many parameters, but it is unclear whether they work in few-shot settings. In this work we present Atlas, a carefully designed and pre-trained retrieval augmented language model able to learn knowledge intensive tasks with very few training examples. We perform evaluations on a wide range of tasks, including MMLU, KILT and NaturalQuestions, and study the impact of the content of the document index, showing that it can easily be updated. Notably, Atlas reaches over 42% accuracy on Natural Questions using only 64 examples, outperforming a 540B parameters model by 3% despite having 50x fewer parameters.