GRASP: GRanularity-Aware Search Policy for Agentic RAG

2026-07-11

AI Summary

This work addresses the challenge of enabling agents to dynamically decide when, how (via semantic or keyword search), and at what contextual granularity to retrieve information during multi-hop reasoning. To this end, the paper proposes GRASP, a novel framework that uniquely integrates contextual granularity control with coordinated multi-tool retrieval. Trained via reinforcement learning, the agent adaptively interleaves semantic search, keyword search, and paragraph reading, achieving a dynamic balance among exploration, verification, and precise retrieval while exhibiting interpretable “browse-and-scan” behavior. A carefully designed joint reward function incorporates answer accuracy, evidence grounding, retrieval complementarity, and turn efficiency. Experiments demonstrate that GRASP significantly outperforms single-step retrieval, prompt-based RAG, and existing reinforcement learning–based retrieval methods on multi-hop reasoning benchmarks, substantially improving both retrieval recall and question-answering accuracy.

Abstract

Agentic retrieval-augmented generation (RAG) extends static RAG by allowing language models to iteratively reason, generate search queries, retrieve evidence, and predict answers. However, it remains challenging for models to decide when to retrieve, whether to use lexical matching or semantic similarity, and how to control context granularity to prevent irrelevant tokens from interfering with agent reasoning. In this paper, we introduce GRASP, a reinforcement learning (RL) framework for training agents to adaptively coordinate complementary retrieval tools during multi-step reasoning. GRASP provides the agent with semantic search, keyword search, and paragraph-reading actions, enabling it to retrieve sentence-level evidence and expand further context only when needed. We train the policy with a reward that jointly accounts for answer accuracy, grounded reading, complementary search, and turn efficiency. Experiments on multi-hop reasoning benchmarks show that GRASP improves both retrieval recall and downstream question answering performance compared with single-step retrieval, prompting-based agentic RAG, and RL-based retrieval baselines. Qualitative and ablation analyses show that the learned policy develops interpretable skimming and scanning behavior: it uses semantic search for broad exploration, paragraph reading for local verification, and keyword search for entity-specific evidence. These results suggest that learning to coordinate retrieval signals and context granularity is critical for agent's correct reasoning.

Problem

• Agentic RAG
• retrieval policy
• context granularity
• multi-step reasoning
• retrieval coordination

Innovation

• granularity-aware retrieval
• agentic RAG
• reinforcement learning
• multi-step reasoning
• adaptive search policy