When Retrieval Doesn't Help: A Large-Scale Study of Biomedical RAG

In three linesLarge-scale study across 5 models (7B-72B parameters), 10 biomedical QA datasets, and 4 retrieval methods. RAG yields only +1-2 point improvements over no-retrieval baseline. Backbone model choice has larger impact than retriever or corpus selection. Bottleneck: model's limited ability to effectively use retrieved evidence.

Summary generated by Claude — human-verified