# Few-shot Learning with Retrieval Augmented Language Model

**Authors:** Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Yu, Armand Joulin, Sebastian Riedel, Edouard Grave  
**arXiv:** [2208.03299](https://arxiv.org/abs/2208.03299)  
**Date:** 2022-08-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.03299v1/

## Key points

- Few-shot learning with retrieval augmented language model
- Large language models in few-shot learning scenarios
- Knowledge-intensive tasks like question answering and fact checking
- Atlas: meticulously designed and pre-trained retrieval augmented language model
- Achieving over 42% accuracy on Natural Questions using just 64 examples
- Outperforming a model with 540 billion parameters by 3%

## Summary

In their paper titled "Few-shot Learning with Retrieval Augmented Language Model," authors Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave explore the capabilities of large language models in few-shot learning scenarios. They highlight that while these models have demonstrated impressive results across various tasks,<kg>few-shot learning</kg>, <kg>retrieval augmented language model</kg>, and <kg>large language models</kg>, knowledge-intensive tasks like question answering and fact checking often require a substantial number of parameters to store relevant information. To address this challenge,<kg>knowledge-intensive tasks</kg>, the authors introduce Atlas,<kg>meticulously designed and pre-trained retrieval augmented language model</kg>, that excels at learning knowledge-intensive tasks with minimal training examples. Through evaluations on tasks such as MMLU,<kg>MMLU</kg>, KILT,<kg>KILT</kg>, and NaturalQuestions,<kg>NaturalQuestions</kg>, they demonstrate Atlas's ability to achieve over 42% accuracy on Natural Questions using just 64 examples.<kf>Remarkably,</kf> Atlas outperforms a model with 540 billion parameters by 3%, despite having significantly fewer parameters (50 times less).