EVENT DOSSIER
A study published on arXiv indicates that large language models often exhibit “epistemological narrowness” when supporting knowledge-intensive tasks. This phenomenon is characterized by the model repeatedly collapsing a large set of valid answers into a small set of normative subsets, showing only some reasoning paths. The study suggests that “epistemological diversity” should be used as an evaluation dimension, defined as the range of valid answers, explanations, and reasoning paths presented by the model to users. Existing evaluation paradigms rely too heavily on accuracy or alignment; this paper argues that we need to go beyond these limitations, treating epistemological diversity as a key indicator of model capabilities, and constructing a preliminary conceptual and measurement framework.
大型语言模型（LLMs）在支持知识密集型任务时，其评估不能仅依赖准确性或对齐度。本文提出“认识论多样性”作为重要评估维度，将其定义为 LLM 向用户展示的有效答案、解释及推理路径的范围。研究构建了概念化与测量该维度的初步框架，并在两个领域进行实证操作。研究发现，前沿 LLMs 常表现出认识论狭隘性，反复将大量有效答案空间坍缩至少量规范子集。这些发现表明，LLM 评估应超越以准确性为导向的范式，将认识论多样性视为模型能力的关键维度。