GRASP: GRanularity-Aware Search Policy for Agentic RAG

GRASP is a reinforcement-learning framework for agentic RAG that coordinates semantic search, keyword search, and paragraph reading during multi-step reasoning. The policy can retrieve sentence-level evidence and expand to broader context only when needed. Its reward jointly considers answer accuracy, grounded reading, complementary retrieval, and turn efficiency. According to the abstract, experiments on multi-hop reasoning benchmarks show improved retrieval recall and downstream QA performance over single-step retrieval, prompting-based agentic RAG, and RL retrieval baselines. Qualitative and ablation analyses reportedly reveal interpretable exploration, verification, and entity-focused search behaviors.