Deepsearch agents improve query focus and evidence quality in searches
QREF: Improving DeepSearch Agents by Aligning Search with Reasoning via Query Refinement and Evidence Filtering
Artificial Intelligence
Summary
DeepSearch agents help computers find complex information by searching and reasoning step-by-step. The authors found that these agents sometimes send unclear or repeated questions, causing less useful results. They created QREF, which rewrites search questions to better match the agent’s thinking and filters out irrelevant information. This makes the search more focused and reduces noise, improving the agent’s overall performance without changing its core design.
What this means in practice
- •For search engine developers: Enable search engines to produce more precise queries and better filter search results based on the current reasoning context, improving retrieval quality.
- •For knowledge management teams: Improve internal research tools to reduce redundant or unclear queries and focus on relevant evidence for complex information tasks.
Authors
Yujia Chen, Yuhui Wang, Junzhe Wang, Shihan Dou, Di Liang, Yujiong Shen, Yuhao Zhou, Zhiheng Xi, Ming Zhang, Senjie Jin, Jiayi Chen, Dingwei Zhu, Yajie Yang, Yuan Hua, Chenyao Lu, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract
DeepSearch agents achieve strong performance on complex information-seeking tasks through iterative search and reasoning. However, they often issue intent-misaligned, under-constrained, or redundant search queries, resulting in poor evidence coverage and accumulating retrieval noise. To mitigate these limitations, we propose QREF, a plug-and-play Query Refiner and Evidence Filter. Specifically, the Query Refiner rewrites search queries according to the current reasoning intent while ensuring complementary coverage across each search batch. The Evidence Filter retains evidence relevant to both the current search batch and the reasoning objective, reducing distracting retrieval results. We initialize the Query Refiner with supervised fine-tuning and optimize it through reinforcement learning based on interactions with a real-world web search environment, reducing ineffective queries. Comprehensive experiments demonstrate that our approach improves the performance of DeepSearch agents across four challenging benchmarks, achieving average relative Pass@1 improvements of 33.55% on WebSailor and 11.04% on TongyiDeepResearch without modifying architectures or retraining their reasoning policies. Compared with the original DeepSearch backbone without QREF, the RL-trained refiner reduces the fraction of query batches that lack discriminative retrieval anchors by 95%. Compared with the SFT-trained refiner, it also lowers the filter-based ineffective query ratio.