Reproduce “GRASP: GRanularity-Aware Search Policy for Agentic RAG” on OpenResearch

About this paper

Varun Gandhi, Jaewook Lee, Shantanu Todmal, Franck Dernoncourt, Ryan Rossi, Zichao Wang, Andrew Lan

GRASP trains an agent to coordinate semantic search, keyword search, and sentence-to-parent-paragraph reading across multi-hop reasoning steps. Its policy chooses how to retrieve evidence and when to expand context using reinforcement learning. On 2WikiMultiHopQA, it reaches passage recall of 0.90, versus 0.83 for IRCoT, and token-level F1 of 0.60, versus 0.53 for Search-R1 (GRPO), in a specific evaluation setup.

Key insights

• The learned GRASP policy exhibits emergent human-like information foraging behaviors, such as using semantic search for broad exploration and keyword search for targeted, entity-specific retrieval.
• Adaptive control over context granularity, allowing agents to retrieve fine-grained sentences and only expand to broader paragraphs when necessary, is crucial for generating precise intermediate queries and effective multi-hop reasoning.
• A comprehensive, multi-objective reward function can effectively train a policy to coordinate diverse retrieval tools and manage context, enabling a smaller 3B-parameter LLM to achieve competitive performance against larger models in complex QA tasks.

Summary by alphaXiv.