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

TL;DR

An adaptive QA framework that dynamically selects retrieval-augmented LLM strategies (no retrieval, single-step retrieval, iterative retrieval) via a classifier predicting question complexity, enhancing efficiency and accuracy.

Problem

Existing retrieval-augmented LLMs apply the same retrieval strategy to all queries, causing unnecessary computational cost for simple queries and inadequate handling of complex multi-step queries. User queries have a continuous complexity spectrum, not a simple binary, requiring adaptive strategies.

Approach

1) Define three complexity levels: simple (no retrieval), moderate (single retrieval), complex (iterative retrieval). 2) Automatic labeling: For each query, execute all three strategies and assign the complexity level of the strategy that yields the most accurate answer. 3) Train a small language model (e.g., BERT) as a classifier to predict query complexity. 4) At inference, select the appropriate strategy based on the predicted complexity.

Results & Contribution

Experiments on multiple open-domain QA datasets with varying query complexities show that the proposed method outperforms single-strategy and existing adaptive retrieval methods in both average accuracy and efficiency (processing time). It achieves a balance by reducing unnecessary retrieval for simple queries and improving accuracy for complex queries through iterative retrieval. Code is publicly available for reproducibility.