GRASP Improves Agentic RAG with Granularity-Aware Search Policy

Agentic Retrieval-Augmented Generation (RAG) systems enhance static RAG by allowing language models to iteratively reason, generate queries, retrieve evidence, and formulate answers. However, these systems face challenges in deciding when to retrieve, which retrieval method to use (lexical vs. semantic), and how to manage context granularity to prevent irrelevant information from interfering with reasoning. To address these issues, researchers introduce GRASP (GRanularity-Aware Search Policy), a reinforcement learning (RL) framework. GRASP trains agents to adaptively coordinate various retrieval tools, such as semantic search for broad exploration, keyword search for entity-specific evidence, and paragraph reading for local verification. This allows the agent to retrieve evidence at a sentence level and expand context only when necessary, optimizing for relevance and efficiency. The policy is trained with a reward function that balances answer accuracy, grounded reading, complementary search, and turn efficiency. Experiments on multi-hop reasoning benchmarks show that GRASP significantly improves both retrieval recall and downstream question answering performance compared to existing baselines. Qualitative analysis indicates that the learned policy develops intelligent "skimming and scanning" behaviors, demonstrating the critical role of learning to coordinate retrieval signals and context granularity for accurate agent reasoning.