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:
- Generating search queries
- Retrieving evidence
- Continuing to reason and provide an answer
The challenge lies in the model needing to learn:
- When to retrieve
- Whether to use lexical or semantic search
- How granular the context should be, avoiding irrelevant tokens that disrupt reasoning
Method
GRASP incorporates three types of actions into the policy space:
- semantic search
- keyword search
- paragraph reading
Training utilizes a joint reward that comprehensively considers:
- Answer accuracy
- Evidence groundedness (grounded reading)
- Search complementarity
- Step efficiency
Experimental Conclusions
On multi-hop reasoning benchmarks, GRASP outperforms the following methods:
- Single-step retrieval
- Prompt-driven agentic RAG
- RL-based retrieval baselines
Behavioral Analysis
The authors also found the learned policy to be highly interpretable:
- Semantic search is better suited for broad exploration
- Paragraph reading is preferred for local verification
- Keyword search is ideal for entity-level evidence localization
Conclusion: Enabling models to coordinate different retrieval signals and context granularities is critical for correct agent reasoning.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21