Large Language Models Struggle to Learn Long-Tail Knowledge

TL;DR

Large language models' ability to learn rare, long-tail knowledge heavily depends on the amount of related information seen during pre-training, and retrieval-augmentation offers a promising solution to improve performance on such knowledge.

Contribution

This paper establishes a strong correlation between pre-training data support and model accuracy on long-tail knowledge, and demonstrates retrieval-augmentation as an effective method to mitigate data sparsity issues.

Findings

Model accuracy correlates with relevant document count in pre-training data.

Larger models perform better on long-tail knowledge but still require massive scaling.

Retrieval-augmentation reduces dependence on pre-training data for rare facts.

Abstract

The Internet contains a wealth of knowledge -- from the birthdays of historical figures to tutorials on how to code -- all of which may be learned by language models. However, while certain pieces of information are ubiquitous on the web, others appear extremely rarely. In this paper, we study the relationship between the knowledge memorized by large language models and the information in pre-training datasets scraped from the web. In particular, we show that a language model's ability to answer a fact-based question relates to how many documents associated with that question were seen during pre-training. We identify these relevant documents by entity linking pre-training datasets and counting documents that contain the same entities as a given question-answer pair. Our results demonstrate strong correlational and causal relationships between accuracy and relevant document count for numerous question answering datasets (e.g., TriviaQA), pre-training corpora (e.g., ROOTS), and model sizes (e.g., 176B parameters). Moreover, while larger models are better at learning long-tail knowledge, we estimate that today's models must be scaled by many orders of magnitude to reach competitive QA performance on questions with little support in the pre-training data. Finally, we show that retrieval-augmentation can reduce the dependence on relevant pre-training information, presenting a promising approach for capturing the long-tail.