GRASP: GRanularity-Aware Search Policy for Agentic RAG

2026-07-11 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

GRASP, a reinforcement learning (RL) framework, trains agents to adaptively coordinate complementary retrieval tools within agentic retrieval-augmented generation (RAG) systems. It addresses challenges in deciding when and how to retrieve information and controlling context granularity to prevent irrelevant tokens from interfering with agent reasoning. GRASP provides agents with semantic search, keyword search, and paragraph-reading actions, enabling retrieval of sentence-level evidence and context expansion only when necessary. The policy is trained using a reward system that jointly accounts for answer accuracy, grounded reading, complementary search, and turn efficiency. Experiments on multi-hop reasoning benchmarks demonstrate that GRASP improves both retrieval recall and downstream question answering performance compared to single-step retrieval and prompting-based baselines. The learned policy exhibits interpretable skimming and scanning behaviors, utilizing semantic search for broad exploration, paragraph reading for local verification, and keyword search for entity-specific evidence.

Key takeaway

For AI Scientists and Machine Learning Engineers developing agentic RAG systems, improving multi-hop reasoning performance requires a sophisticated approach to retrieval. You should prioritize implementing adaptive policies that coordinate diverse retrieval tools like semantic and keyword search with dynamic context granularity. Consider using reinforcement learning frameworks, such as GRASP, to train agents that can intelligently skim and scan information, optimizing both recall and question answering accuracy in complex reasoning tasks.

Key insights

Learning to coordinate diverse retrieval signals and context granularity is critical for robust agentic reasoning.

Principles

• Adaptive coordination of complementary retrieval tools enhances agentic RAG.
• Dynamic context granularity prevents irrelevant token interference.
• Multi-objective reward functions improve agent training efficiency.

Method

GRASP trains RL agents to adaptively coordinate semantic search, keyword search, and paragraph-reading actions during multi-step reasoning, retrieving sentence-level evidence and expanding context only as needed.

In practice

• Integrate diverse search actions (semantic, keyword, reading) into RAG agents.
• Develop RL-based policies for dynamic context management.
• Design reward signals balancing accuracy, grounding, and efficiency.

Topics

• Agentic RAG
• Reinforcement Learning
• Information Retrieval
• Semantic Search
• Keyword Search
• Multi-hop Reasoning
• Context Granularity

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Agentic RAG: Let the Agent Search — Agentic RAG empowers LLMs to iteratively search, read, and decide on evidence, overcoming static RAG limitations.
• Agentic RAG-VLM: Affordance-Aware Retrieval-Augmented Generation with Self-Reflective Planning for Robotic Grasping — Agentic RAG-VLM enhances robotic grasping by integrating affordance-aware RAG, scene graph reasoning, and self-reflective planning.
• RAG Has Evolved into Multi-modal AI Search Agents (RL, HaPO) — Multimodal AI search agents utilize knowledge graphs and checkpoint-based RL for efficient multi-step visual reasoning.
• The Real Leap Is Not Better Retrieval. It’s Better Control Over Retrieval. — Agentic RAG shifts focus from static retrieval to dynamic control over the search process.
• Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search — Slim improves long-horizon agentic search by managing context efficiently through distinct search/browse tools and periodic summarization.
• Closing the Retriever Gap in Agentic Search Systems, Offline Negative Item Filtering at Scale, and More! — Recent advances enhance RAG and recommender systems through hierarchical retrieval, negative feedback modeling, and interpretable bias control.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.