# Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs

**Authors:** Oded Ovadia, Menachem Brief, Moshik Mishaeli, Oren Elisha  
**arXiv:** [2312.05934](https://arxiv.org/abs/2312.05934)  
**Date:** 2023-12-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.05934v3/

## Key points

- Study conducted by Oded Ovadia, Menachem Brief, Moshik Mishaeli, and Oren Elisha
- Explores effectiveness of unsupervised fine-tuning and retrieval-augmented generation (RAG) in enhancing Large Language Models (LLMs)
- LLMs contain vast factual information but limited by training data characteristics
- RAG consistently outperforms unsupervised fine-tuning in incorporating existing and new knowledge
- LLMs struggle to learn new factual information through unsupervised fine-tuning
- Data augmentation can enhance language model performance and address challenges like Berglund et al.'s Reversal Curse
- Importance of further research on how large language models adapt to new knowledge
- RAG emerges as a more reliable choice for effective knowledge injection compared to traditional fine-tuning methods

## Summary

In the study "Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs," conducted by Oded Ovadia, Menachem Brief, Moshik Mishaeli, and Oren Elisha, the researchers explore the effectiveness of two common approaches - unsupervised fine-tuning and retrieval-augmented generation (RAG) - in enhancing the capabilities of Large Language Models (LLMs). LLMs are known to contain a vast amount of factual information within their pre-trained weights, allowing them to answer diverse questions across different domains. However, this knowledge is limited by the characteristics of the training data. The researchers evaluated both approaches on various knowledge-intensive tasks spanning different topics. Their findings indicate that while unsupervised fine-tuning offers some improvement, RAG consistently outperforms it in incorporating existing knowledge encountered during training as well as entirely new knowledge. The study also highlights that LLMs struggle to learn new factual information through unsupervised fine-tuning and suggests that exposing them to multiple variations of the same fact during training could alleviate this issue. Furthermore, the research delves into data augmentation as a method to enhance language model performance and discusses its potential benefits in increasing LM,Q and addressing challenges like Berglund et al. 's Reversal Curse. The study concludes by emphasizing the importance of exploring further research avenues in understanding how large language models adapt to new knowledge, with RAG emerging as a more reliable choice for effective knowledge injection compared to traditional fine-tuning methods. This work contributes valuable insights into optimizing LLM performance and adapting them to incorporate new information effectively.