Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs
This addresses the challenge of updating knowledge in LLMs for AI practitioners, but it is incremental as it compares existing methods.
The study compared unsupervised fine-tuning and retrieval-augmented generation (RAG) for injecting knowledge into large language models, finding that RAG consistently outperforms fine-tuning on knowledge-intensive tasks, with fine-tuning struggling to learn new facts.
Large language models (LLMs) encapsulate a vast amount of factual information within their pre-trained weights, as evidenced by their ability to answer diverse questions across different domains. However, this knowledge is inherently limited, relying heavily on the characteristics of the training data. Consequently, using external datasets to incorporate new information or refine the capabilities of LLMs on previously seen information poses a significant challenge. In this study, we compare two common approaches: unsupervised fine-tuning and retrieval-augmented generation (RAG). We evaluate both approaches on a variety of knowledge-intensive tasks across different topics. Our findings reveal that while unsupervised fine-tuning offers some improvement, RAG consistently outperforms it, both for existing knowledge encountered during training and entirely new knowledge. Moreover, we find that LLMs struggle to learn new factual information through unsupervised fine-tuning, and that exposing them to numerous variations of the same fact during training could alleviate this problem.