Self-RAG

“Adaptive methods such as FLARE~jiang2023active, Self-RAG~asai2024selfrag, and DRAGIN~su2024dragin dynamically trigger retrieval based on uncertainty signals, but do so reactively, first detecting uncertainty and then blocking generation to perform retrieval.”

“However, their reliance on the LLM itself or the training data makes their generalization susceptible to data biases.”

“Prior attempts like Self-RAG introduce special tokens to control reasoning but require architectural modifications.”

“However, these methods still operate at the document level, failing to adequately filter individual text chunks.”

“this solution requires the training of two external models, requiring tens of thousands of additional training samples”

“these methods typically incur high inference latency due to multiple LLM calls”

“While effective in identifying valuable documents, multiple LLM calls introduce substantial computation overhead.”

“Self-RAG requires threshold tuning to balance QA performance and retrieval efficiency, while vanilla prompting is insufficient in guiding LLMs to make reliable retrieval decisions”

“While these methods improve robustness against irrelevant context, they typically operate via Breadth-First Addition: they append new passages to the existing context.”

“However, these methods generally require substantial computational resources and API costs, making model updates challenging.”

“While effective, these approaches often add supervision, special control tokens, auxiliary probers, or multi-stage loops that increase engineering complexity and latency.”