Retrieval-Augmented Reasoning for Knowledge-Intensive AI Systems Using Dynamic Evidence Acquisition

Abstract

Retrieval-Augmented Generation (RAG) improves factual grounding by adding external evidence to large language model prompts, but conventional pipelines usually retrieve evidence only once before generation. This static design is limited for knowledge-intensive tasks in which intermediate reasoning steps expose new evidence needs. This paper presents Retrieval-Augmented Reasoning (RAR), an architecture that embeds retrieval inside the reasoning workflow using a reasoning-state manager, retrieval-trigger decisions, evidence integration, and answer generation. The system is evaluated using a controlled 25-scenario multi-hop workload, a deterministic $\text{1, 3 0 0}$-scenario stress test, and a public-validation workflow based on HotpotQA-style and RAGTruth-style examples. In the controlled workload, RAR Stateful achieves the strongest F1 balance (0.614 +/- 0.136) and improves answer coverage over Static RAG Top-3 from 0.589 to 0.747. In the stress test, RAR Stateful reaches 60.77% F1 success and 76.23% unsupported-control success. Public validation shows 90.0% evidence recall and grounded-answer rate on HotpotQA-style examples and 85.0% unsupported-response detection, fallback correctness, and safe-consistency on RAGTruth-style examples. The results show that reasoning-guided retrieval improves answer completeness and reliability with manageable retrieval-control overhead.