GRASP: Graph Agentic Search for Multi-hop QA
arXiv:2605.16598 introduces GRASP (Graph Agentic Search over Propositions), an agentic retrieval system for multi-hop question answering. It optimizes both accuracy and token usage by decomposing queries into dependency-aware plans and dynamically scaling sub-agents. Each sub-agent explores a three-layer hierarchical graph of entities, propositions, and passages. The system avoids expensive graph construction and reduces inference token usage compared to prior methods.
Key facts
- GRASP stands for Graph Agentic Search over Propositions.
- It is designed for multi-hop question answering.
- It decomposes queries into dependency-aware plans.
- It dynamically scales sub-agents based on problem complexity.
- It uses a three-layer hierarchical graph of entities, propositions, and passages.
- It aims to reduce token usage at inference time.
- It avoids expensive graph construction at index time.
- The paper is on arXiv with ID 2605.16598.
Entities
Institutions
- arXiv