DRAGIN

“These signals can offer insights into the model's reasoning process, but often require access to model weights and, consequently, cannot be used with closed-source LLMs. Additionally, the lack of interpretability in latent representations undermines trustworthiness, making these signals less suitable for applications such as healthcare, where trustworthiness plays a critical role.”

“This demonstrates that relying on a single token's confidence as the signal can result in ineffective intervention, allowing multiple low-confidence tokens to be generated before retrieve.”

“However, current dynamic RAG methods fail to predict whether the LLM has the capability to answer a question prior to generation, thereby triggering retrieval in advance. Moreover, most methods often rely on static rules, leading to ineffective timing for retrieval triggers during the generation process.”

“However, the former approach has limitations as LLMs tend to be overconfident, generating high-confidence probability distributions even when lacking relevant knowledge.”