DeepSearch agents are designed to handle complex information-seeking tasks by alternating between search and reasoning. However, the paper notes that these agents frequently generate queries that are poorly aligned with their actual intent, too broad, or redundant. Such queries can undermine the effectiveness of the iterative process.
The authors propose QREF, a method that explicitly aligns search actions with the agent's reasoning state. It does so by refining queries before they are issued and filtering evidence after retrieval, presumably to keep only the most relevant information for the next reasoning step.
The abstract does not include experimental results, so the claimed benefits are not quantified here. Still, the work points to a specific, practical weakness in current DeepSearch designs and offers a targeted remedy. No other sources were available for comparison, so this summary reflects only the arXiv preprint.