DeepSearch agents achieve strong performance on complex information‑seeking tasks through iterative search and reasoning, yet they often issue intent‑misaligned, under‑constrained, or redundant queries, resulting in poor evidence coverage and accumulating retrieval noise. To address these issues we introduce QREF, a plug‑and‑play Query Refiner and Evidence Filter. The Query Refiner rewrites search queries according to the current reasoning intent and ensures complementary coverage across each query batch; the Evidence Filter retains evidence that is relevant to both the current batch and the reasoning objective, reducing distracting results. The refiner is first initialized with supervised fine‑tuning and then optimized via reinforcement learning through interactions with a real‑world web search environment, markedly decreasing ineffective queries. Comprehensive experiments on four challenging benchmarks show average relative Pass@1 improvements of 33.55% on WebSailor and 11.04% on TongyiDeepResearch, without altering architectures or retraining reasoning policies. Compared with the original DeepSearch backbone without QREF, the RL‑trained refiner cuts the fraction of query batches lacking discriminative retrieval anchors by 95% and also lowers the filter‑based ineffective query ratio relative to the SFT‑trained refiner.
Review: QREF tightly aligns search with reasoning by dynamically refining queries and filtering evidence, substantially boosting retrieval quality and task success rates, and offering a practical pathway toward more efficient intelligent search systems.