GRASP Teaches Agentic RAG When to Search

Varun Gandhi · hf · 2026-07-17

Research Content

This paper introduces GRASP, a method for training the retrieval strategy of Agentic RAG. Unlike static RAG, Agentic RAG allows the model to repeatedly perform multi-step reasoning:

The challenge lies in the model needing to learn:

Method

GRASP incorporates three types of actions into the policy space:

Training utilizes a joint reward that comprehensively considers:

Experimental Conclusions

On multi-hop reasoning benchmarks, GRASP outperforms the following methods:

Behavioral Analysis

The authors also found the learned policy to be highly interpretable:

Conclusion: Enabling models to coordinate different retrieval signals and context granularities is critical for correct agent reasoning.

Original post →

More from coding & agent

coding & agent channel →