Only Ask What You Don't Know: Grounded Delta Planning for Efficient Multi-step RAG

The article presents Grounded Delta Planning RAG (GDP-RAG), a novel framework for multi-hop question answering that enhances Retrieval-Augmented Generation (RAG) by focusing on information deltas. Key innovations include preliminary retrieval for grounding, a gap-conditioned planning prompt to target missing information, and a skeletal trajectory for subqueries, which collectively improve accuracy to 60.63% on benchmarks like HotpotQA while reducing computational costs significantly compared to existing methods. This approach is crucial for practitioners as it optimizes resource usage while enhancing the reliability of multi-step reasoning in AI applications.