📄 中文摘要
在提升生成模型的可信性和稳健性方面，不确定性量化（UQ）方法展现出良好的潜力。然而，许多现有方法依赖于僵化的启发式规则，难以在不同任务和模态中推广。提出了一种新颖的UQ框架，具有高度灵活性，能够接近或超越先前启发式方法的性能。引入了方向性集中不确定性（DCU），这是一种基于冯·米塞斯-费舍尔（vMF）分布的统计程序，用于量化嵌入的集中度。该方法通过测量语言模型生成的多个输出的几何分散性，捕捉不确定性。
方向性集中不确定性：一种针对生成模型的不确定性量化的表征方法
📄 English Summary
Directional Concentration Uncertainty: A representational approach to uncertainty quantification for generative models
This research proposes a novel framework for Uncertainty Quantification (UQ) that demonstrates high flexibility and approaches or surpasses the performance of previous heuristic methods. It introduces Directional Concentration Uncertainty (DCU), a new statistical procedure for quantifying the concentration of embeddings based on the von Mises-Fisher (vMF) distribution. The method captures uncertainty by measuring the geometric dispersion of multiple generated outputs from a language model using continuous embeddings of the generated outputs. This approach addresses the limitations of existing UQ methods that rely on rigid heuristics, making it applicable across various tasks and modalities.