Baidu's Improving Retrieval Augmented Language Model with Self-Reasoning
arxiv.orgCan anyone explain what is gained by training a model? Why not use the foundational LLM for the relevance, evidence, and trajectory processes?
I assume you are referring to fine tuning a model here?
You could also just continue pre-training of an existing foundation model. Would still be cheaper by not starting from zero.
The amount of accuracy while doing fine tuning or distillation is usually better than pre-training an existing model, not to mention the graph against the cost.