LLM-Independent Adaptive RAG: Let the Question Speak for Itself

Large Language Models (LLMs) are prone to hallucinations, and Retrieval-Augmented Generation (RAG) helps mitigate this, but at a high computational cost while risking misinformation. Adaptive retrieval aims to retrieve only when necessary, but existing approaches rely on LLM-based uncertainty estimation, which remain inefficient and impractical. In this study, we introduce lightweight LLM-independent adaptive retrieval methods based on external information. We investigated 27 features, organized into 7 groups, and their hybrid combinations. We evaluated these methods on 6 QA datasets, assessing the QA performance and efficiency. The results show that our approach matches the performance of complex LLM-based methods while achieving significant efficiency gains, demonstrating the potential of external information for adaptive retrieval.

Introduction. Large Language Models (LLMs) excel in tasks like question answering (QA) (Yang et al., 2018; Kwiatkowski et al., 2019), but remain vulnerable to hallucinations (Yin et al., 2024; Ding et al., 2024). Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) mitigates this by incorporating external information, although it introduces risks such as error accumulation (Shi et al., 2023) and external hallucinations (Ding et al., 2024). Adaptive retrieval techniques (Moskvoretskii et al., 2025; Ding et al., 2024; Jeong et al., 2024) aim to balance LLM knowledge with external resources by estimating uncertainty to decide whether retrieval is needed. However, existing methods primarily frame this task as uncertainty estimation based on LLM internal states or outputs, leading to significant computational overhead. This can offset the efficiency gains from reduced retrieval calls and limit practicality, especially with larger models.

Discussion / Conclusion. In this work, we introduced 7 groups of lightweight external features for LLM-independent adaptive retrieval, improving efficiency by eliminating the need for LLM-based uncertainty estimation while preserving QA performance. Our approach outperforms uncertainty-based methods for complex questions and offers a detailed analysis of the complementarity between uncertainty and external features.