📄 中文摘要
用户越来越依赖大型语言模型（LLMs）进行深度研究，利用它们将多样化的来源综合成结构化报告，以支持理解和行动。在此背景下，这些报告的实际可靠性取决于逻辑质量，即报告中的主张和论点是否得到明确支持，并能被信任作为后续使用的基础，而不仅仅是看起来流畅或信息丰富。然而，目前的评估框架在很大程度上忽视了这一要求。为填补这一空白，提出了ReportLogic，一个通过读者中心的可审计性视角量化报告级逻辑质量的基准。具体而言，ReportLogic采用了一个分层分类法，评估读者是否能够...
ReportLogic：评估深度研究报告的逻辑质量
出处: ReportLogic: Evaluating Logical Quality in Deep Research Reports
发布: 2026年2月24日
📄 English Summary
ReportLogic: Evaluating Logical Quality in Deep Research Reports
Users increasingly rely on Large Language Models (LLMs) for Deep Research, synthesizing diverse sources into structured reports that facilitate understanding and action. The practical reliability of these reports hinges on their logical quality, which concerns whether claims and arguments are explicitly supported and can be trusted for downstream use, rather than merely appearing fluent or informative. Current evaluation frameworks largely overlook this requirement. To address this gap, ReportLogic is introduced as a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy to evaluate whether readers can...