New AI model cuts retrieval calls by 3.7% while preserving accuracy

Researcher Weimeng Luo has published a paper detailing an optimization for multi-round retrieval-augmented generation (RAG) systems that improves efficiency without significantly compromising accuracy. The work focuses on the critical decision of when to stop retrieving additional evidence during multi-round RAG processes, which the author frames as a sequential selection problem rather than an independent classification task.

The solution adapts the S2G-RAG structured sufficiency-and-gap judgment approach to a frozen Search-R1 pipeline. A Qwen3.5-2B model was trained on 3,009 states derived from 900 HotpotQA questions to make better stopping judgments. During evaluation on a confirmatory test set, the trained policy reduced retrieval calls by 77 instances (3.70%) compared to the native Search-R1 baseline, while only decreasing exact match accuracy by 0.625 percentage points. This suggests that judicious stopping decisions can significantly reduce computational overhead in RAG systems with minimal impact on answer quality.