Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity

by Korea Advanced Institute of Science and Technology (KAIST) · free · Last verified 2026-03-17

Proposes Adaptive-RAG, a framework that learns to select the most suitable retrieval strategy for each question based on its complexity using a small classifier. The approach dynamically routes queries to no-retrieval, single-step, or multi-step retrieval strategies, balancing accuracy and efficiency across question types.