Adaptive-RAG: Dynamically Selecting Retrieval Strategies Based on Question Complexity

Recent advances in Large Language Models (LLMs) have led to substantial improvements across a wide range of natural language processing tasks, particularly in open-domain Question Answering (QA). Models such as GPT-style architectures and instruction-tuned variants including the FLAN-T5 (Brown et al., 2020; Chung et al., 2022) family demonstrate strong language understanding and reasoning capabilities, enabling them to generate fluent and contextually relevant answers to a diverse set of factual, descriptive, and explanatory questions .These models benefit from large-scale pretraining on massive text corpora, which allows them to capture rich linguistic patterns and a broad range of world knowledge.
Despite these notable successes, LLMs remain fundamentally constrained by their reliance on parametric knowledge acquired during pretraining. Their internal knowledge is static and cannot be easily updated, which makes them prone to generating factually incorrect, incomplete, or outdated answers when responding to knowledge-intensive or time-sensitive queries. Furthermore, LLMs may hallucinate plausible-sounding information when they lack sufficient knowledge about a topic, reducing their reliability in real-world QA applications.
To mitigate these limitations, Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) has emerged as an effective approach for improving the factual accuracy and robustness of LLM-based question answering systems. RAG enhances LLMs by incorporating external, non-parametric knowledge retrieved from large document collections such as Wikipedia or web-scale corpora (Lewis et al., 2020; Wikipedia, 2025). By conditioning answer generation on retrieved evidence, retrieval-augmented models can ground their responses in relevant and up-to-date information, thereby reducing hallucinations and improving answer correctness.To mitigate these limitations, Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) has emerged as an effective approach for improving the factual accuracy and robustness of LLM-based question answering systems. RAG enhances LLMs by incorporating external, non-parametric knowledge retrieved from large document collections such as Wikipedia or web-scale corpora (Lewis et al., 2020; Wikipedia, 2025). By conditioning answer generation on retrieved evidence, retrieval-augmented models can ground their responses in relevant and up-to-date information, thereby reducing hallucinations and improving answer correctness.
However, while retrieval-augmented methods have demonstrated significant performance gains, most existing RAG systems employ a fixed retrieval strategy (Lewis et al., 2020; Karpukhin et al., 2020), applying the same retrieval mechanism to all user queries regardless of their difficulty or reasoning requirements. In practical settings, user queries exhibit substantial variation in complexity. Some questions are simple and can be answered accurately using the LLM’s internal knowledge alone, while others require retrieving and reasoning over multiple documents to synthesize an answer.
Applying complex retrieval and multi-step reasoning strategies uniformly to all queries introduces unnecessary computational overhead for straightforward questions. Conversely, relying on minimal or single step retrieval is insufficient for complex multi-hop queries that demand deeper reasoning. These observations highlight the limitations of one-size-fits-all QA systems and motivate the need for an adaptive QA framework that can dynamically select the most appropriate retrieval strategy ranging from no retrieval to multi-step retrieval based on the predicted complexity of each query.