Comparing Knowledge Injection Methods for LLMs in a Low-Resource Regime

TL;DR

This paper investigates methods for injecting small amounts of new knowledge into large language models, comparing augmentation techniques, and analyzing the balance between learning new facts and forgetting, with practical insights and code resources.

Contribution

It introduces diverse augmentation algorithms for knowledge injection in LLMs, analyzes the effects of variability and forgetting, and demonstrates self-generated synthetic data as a promising approach.

Findings

Diverse prompting improves knowledge acquisition more than simple continued pre-training.

Exposing models to varied textual prompts enhances learning of new facts.

RAG-based methods can cause greater degradation on control datasets.

Abstract

Large language models (LLMs) often require vast amounts of text to effectively acquire new knowledge. While continuing pre-training on large corpora or employing retrieval-augmented generation (RAG) has proven successful, updating an LLM with only a few thousand or million tokens remains challenging. In this work, we investigate the task of injecting small, unstructured information into LLMs and its relation to the catastrophic forgetting phenomenon. We use a dataset of recent news -- ensuring no overlap with the model's pre-training data -- to evaluate the knowledge acquisition by probing the model with question-answer pairs related the learned information. Starting from a continued pre-training baseline, we explored different augmentation algorithms to generate synthetic data to improve the knowledge acquisition capabilities. Our experiments show that simply continuing pre-training on limited data yields modest improvements, whereas exposing the model to diverse textual variations significantly improves the learning of new facts -- particularly with methods that induce greater variability through diverse prompting. Furthermore, we shed light on the forgetting phenomenon in small-data regimes, illustrating the delicate balance between learning new content and retaining existing capabilities. We also confirm the sensitivity of RAG-based approaches for knowledge injection, which often lead to greater degradation on control datasets compared to parametric methods. Finally, we demonstrate that models can generate effective synthetic training data themselves, suggesting a pathway toward self-improving model updates. All code and generated data used in our experiments are publicly available, providing a resource for studying efficient knowledge injection in LLMs with limited data at https://github.com/hugoabonizio/knowledge-injection-methods.