GRASP: GRanularity-Aware Search Policy for Agentic RAG
GRASP trains a RAG agent to choose search type and context size during multi-step reasoning.
The paper gives the agent semantic search, keyword search, and paragraph-reading actions, then trains a policy with reinforcement learning. Its reward combines answer accuracy, grounded reading, complementary search use, and turn efficiency. On multi-hop reasoning benchmarks, GRASP beats single-step retrieval, prompted agentic RAG, and RL retrieval baselines in retrieval recall and QA performance. The authors say the learned policy separates broad semantic exploration, local paragraph verification, and entity-specific keyword lookup. HF Daily Papers' note
The paper gives the agent semantic search, keyword search, and paragraph-reading actions, then trains a policy with reinforcement learning. Its reward combines answer accuracy, grounded reading, complementary search use, and turn efficiency. On multi-hop reasoning benchmarks, GRASP beats single-step retrieval, prompted agentic RAG, and RL retrieval baselines in retrieval recall and QA performance. The authors say the learned policy separates broad semantic exploration, local paragraph verification, and entity-specific keyword lookup. HF Daily Papers' note
score 5