Many real-world decisions rely on information search, where people sample evidence and decide when to stop under uncertainty. The uncertainty in the environment, particularly how diagnostic evidence is distributed, causes complexities in information search, further leading to suboptimal decision-making outcomes. Yet AI decision support often targets outcome optimization, and less is known about how to scaffold search without increasing cognitive load. We introduce SERA, an LLM-based assistant that provides either gist or verbatim feedback during search. Across two experiments (N1=54, N2=54), we examined decision-making outcomes and information search in SERA-Gist, SERA-Verbatim, and a no-feedback baseline across three environments varying in uncertainty. The uncertainty in environment is operationalized by the perceived gain of information across the course of sampling, which individuals may experience diminishing return of information gain (decremental; low-uncertainty), or a local drop of information gain (local optimum; medium-uncertainty), or no patterns in information gain (high-uncertainty), as they search more. Individuals show more accurate decision outcomes and are more confident with SERA support, especially under higher uncertainty. Gist feedback was associated with more efficient integration and showed a descriptive pattern of reduced oversampling, while verbatim feedback promoted more extensive exploration. These findings establish feedback representation as a design lever when search matters, motivating adaptive systems that match feedback granularity to uncertainty.

翻译：现实世界中的许多决策依赖于信息搜索过程，即人们在不确定性环境下采集证据并决定何时终止搜索。环境中的不确定性——特别是诊断性证据的分布特征——导致信息搜索过程复杂化，进而引发次优的决策结果。然而现有AI决策支持系统往往聚焦于结果优化，对于如何在不增加认知负荷的前提下构建搜索支架仍知之甚少。本文提出SERA——一种基于大语言模型的辅助系统，能在搜索过程中提供要点式或逐字式反馈。通过两项实验（N1=54，N2=54），我们在三种不同不确定性的环境中，对比了SERA-要点反馈、SERA-逐字反馈及无反馈基线条件下的决策结果与信息搜索行为。环境不确定性通过采样过程中感知信息增益的变化来表征：随着搜索深入，个体可能经历信息增益的边际递减（递减型；低不确定性），或信息增益的局部下降（局部最优；中不确定性），或信息增益的无规律波动（高不确定性）。研究发现：在SERA支持下，个体决策准确率更高且信心更强，这种效应在较高不确定性环境中尤为显著。要点反馈与更高效的信息整合相关，并呈现减少过度采样的趋势性特征；而逐字反馈则促进更广泛的探索行为。这些发现确立了反馈表征作为搜索关键环节的设计杠杆，为开发能根据不确定性匹配反馈粒度的自适应系统提供了理论依据。