RAG vs Fine-Tuning Economics Statistics

Fine-tuning alone improved domain-task accuracy by over 6 percentage points, and RAG alone by about 5 percentage points, in a Microsoft case study

The study evaluated Llama2-13B, GPT-3.5, and GPT-4. RAG captured most of the accuracy gain of the more expensive fine-tuning route, an economically relevant finding for build-versus-retrieve decisions.

Combining RAG and fine-tuning produced cumulative accuracy gains beyond either approach alone

The two techniques addressed different failure modes (knowledge retrieval versus behavioral adaptation), so their gains added rather than overlapped, supporting hybrid pipelines where budget allows.

Retrieval-augmented generation reduces but does not eliminate hallucination, with grounded systems still producing unsupported claims

The survey documents that grounding answers in retrieved context lowers fabrication yet leaves residual faithfulness errors, including citing sources that do not support the claim, which matters acutely in finance.

Domain-adapted financial models such as BloombergGPT outperformed comparable general models on finance tasks, showing the upper bound of pre-training adaptation

BloombergGPT represents the heaviest end of the adaptation spectrum (domain pre-training), more costly than fine-tuning or RAG, and improved finance accuracy without solving general reasoning.

FinBen added a dedicated RAG evaluation track, reflecting that retrieval is now a standard component of financial LLM systems

Including a RAG track in a finance benchmark signals that grounded generation, not raw model recall, is the assumed deployment pattern for finance tasks.