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

Researchers introduce GDP-RAG, a novel retrieval-augmented generation framework that improves multi-hop question answering by focusing computation only on information gaps rather than over-generating reasoning steps. The system achieves 60.63% accuracy on benchmark datasets while reducing computational costs by 22-68% compared to existing approaches.

GDP-RAG addresses a fundamental inefficiency in current retrieval-augmented generation systems: the tendency to either propagate errors through multiple retrieval rounds or wastefully generate excessive reasoning steps that increase costs without improving accuracy. The framework's core innovation lies in its three-part approach—preliminary retrieval for grounding, gap-conditioned planning that targets only missing information, and skeletal trajectories that maintain evidence continuity. This represents a meaningful advancement in RAG efficiency, a critical concern as language models become more expensive to operate at scale.

The research builds on growing recognition that multi-hop question answering requires smarter planning rather than brute-force retrieval. Previous systems like PAR-RAG and KnowTrace attempted to solve this problem but either sacrificed accuracy for cost savings or vice versa. GDP-RAG's dual achievement—highest accuracy among tested systems while maintaining significantly lower computational costs—suggests the field is moving toward more intelligent resource allocation in AI systems.

For developers and organizations deploying RAG systems, this work has immediate practical implications. The 22% cost reduction compared to PAR-RAG and 68% reduction compared to KnowTrace directly impact operational expenses for question-answering applications, search engines, and knowledge-intensive AI systems. This efficiency gain becomes increasingly important as enterprises scale RAG deployments across larger knowledge bases and user bases.

The validation across three major benchmarks (HotpotQA, 2WikiMultiHopQA, MuSiQue) demonstrates robustness across different question-answering domains. Future research will likely explore whether these principles apply to other multi-step reasoning tasks beyond question answering, and whether similar gap-focused planning approaches can optimize other aspects of generative AI systems.

• →GDP-RAG achieves 60.63% accuracy while reducing costs by 22-68% compared to competing systems by focusing only on information gaps.
• →The framework uses preliminary retrieval and gap-conditioned planning to avoid error propagation and unnecessary computation in multi-hop reasoning.
• →Skeletal trajectories pair subqueries with evidence from retrieval, maintaining context throughout the reasoning process.
• →Results are validated across three major benchmarks, demonstrating consistent improvements in both accuracy and efficiency.
• →The approach addresses a critical bottleneck in RAG deployment by balancing quality with computational cost reduction.