# Fine Tuning vs. Retrieval Augmented Generation for Less Popular Knowledge

**Authors:** Heydar Soudani, Evangelos Kanoulas, Faegheh Hasibi  
**arXiv:** [2403.01432](https://arxiv.org/abs/2403.01432)  
**Date:** 2024-03-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.01432v2/

## Key points

- Authors Heydar Soudani, Evangelos Kanoulas, and Faegheh Hasibi discuss challenges faced by large language models (LLMs) with less-popular concepts/entities in domain-specific applications
- Two main approaches highlighted for improving LLM performance: Retrieval Augmented Generation (RAG) and fine-tuning (FT) over synthetic data
- Study evaluates impact of RAG and FT on customizing LLMs for question answering tasks focusing on low-frequency entities
- Findings show FT significantly improves performance across entities of varying popularity, while RAG outperforms other methods
- Success of both RAG and FT attributed to advancements in retrieval and data augmentation techniques
- Importance emphasized of tailored approaches like RAG and FT in optimizing LLMs for handling less-popular knowledge effectively
- Release of data and code on GitHub provides valuable resource for further exploration and development

## Summary

In their paper titled "Fine Tuning vs. Retrieval Augmented Generation for Less Popular Knowledge," authors Heydar Soudani, Evangelos Kanoulas, and Faegheh Hasibi delve into the challenges faced by large language models (LLMs) when dealing with less-popular or low-frequency concepts and entities in domain-specific applications. They highlight two prominent approaches to enhance LLM performance in handling low-frequent topics: Retrieval Augmented Generation (RAG) and fine-tuning (FT) over synthetic data. The study evaluates the impact of RAG and FT on customizing LLMs for question answering tasks specifically focusing on low-frequency entities. The findings reveal that FT significantly improves performance across entities of varying popularity, particularly in the most and least popular groups. On the other hand, RAG outperforms other methods in this context. Furthermore, the success of both RAG and FT approaches is attributed to advancements in retrieval and data augmentation techniques. The authors emphasize the importance of these techniques in optimizing LLMs for handling less-popular knowledge effectively. Overall, this research sheds light on the significance of tailored approaches such as RAG and FT in enhancing LLM performance on low-frequency topics. The release of their data and code on GitHub provides a valuable resource for further exploration and development in this area.