Data, Retrieval & Knowledge
Retrieval-augmented generation outperforms unsupervised fine-tuning for adding knowledge to large language models, study finds
In a 2024 paper at the Conference on Empirical Methods in Natural Language Processing (EMNLP), researchers compared two ways to give a large language model new factual knowledge. Unsupervised fine-tuning means further training the model on data. Retrieval-augmented generation, or RAG, means letting the model consult an outside dataset while it answers. The study found that RAG consistently outperformed unsupervised fine-tuning, both for facts the model had met during training and for wholly new facts. The authors also found that the models had trouble learning new facts through unsupervised fine-tuning alone, but seeing many variations of the same fact during training could help.
What did the study compare?
The work compared unsupervised fine-tuning and retrieval-augmented generation on a variety of knowledge-intensive tasks across different topics. The researchers looked at how well each method handled existing knowledge and entirely new knowledge.
What did the results show?
Unsupervised fine-tuning gave some improvement. Retrieval-augmented generation did better consistently, for knowledge the model had already encountered and for new knowledge. The paper also reports that large language models struggle to learn new factual information through unsupervised fine-tuning, and that showing them many variations of the same fact during training could ease that problem.
Why does exposing the model to variations of the same fact matter?
The finding points to a possible way to help models absorb new facts through fine-tuning. The abstract keeps this as a possibility, not a proven cure.
Limits and what's next
The abstract describes an evaluation on knowledge-intensive tasks across different topics. It does not name the specific large language models, external datasets, or number of tasks. That makes it hard to tell how far the result reaches beyond the setups the authors used. The abstract does not report independent replication, so this is one study. The possible benefit from varied examples is stated as a possibility, not an established remedy.
Source: Oded Ovadia et al., “Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs,” Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024.