Xu Jie

Wenbo Pan, Jie Xu, Qiguang Chen et al.

Large Language Models (LLMs) should refuse to answer questions beyond their knowledge. This capability, which we term knowledge-aware refusal, is crucial for factual reliability. However, existing metrics fail to faithfully measure this ability. On the one hand, simple refusal-based metrics are biased by refusal rates and yield inconsistent scores when models exhibit different refusal tendencies. On the other hand, existing calibration metrics are proxy-based, capturing the performance of auxiliary calibration processes rather than the model's actual refusal behavior. In this work, we propose the Refusal Index (RI), a principled metric that measures how accurately LLMs refuse questions they do not know. We define RI as Spearman's rank correlation between refusal probability and error probability. To make RI practically measurable, we design a lightweight two-pass evaluation method that efficiently estimates RI from observed refusal rates across two standard evaluation runs. Extensive experiments across 16 models and 5 datasets demonstrate that RI accurately quantifies a model's intrinsic knowledge-aware refusal capability in factual tasks. Notably, RI remains stable across different refusal rates and provides consistent model rankings independent of a model's overall accuracy and refusal rates. More importantly, RI provides insight into an important but previously overlooked aspect of LLM factuality: while LLMs achieve high accuracy on factual tasks, their refusal behavior can be unreliable and fragile. This finding highlights the need to complement traditional accuracy metrics with the Refusal Index for comprehensive factuality evaluation.

Ruiqing Ding, Yu Zhou, Jie Xu et al.

The utilization of computer technology to solve problems in medical scenarios has attracted considerable attention in recent years, which still has great potential and space for exploration. Among them, machine learning has been widely used in the prediction, diagnosis and even treatment of Sepsis. However, state-of-the-art methods require large amounts of labeled medical data for supervised learning. In real-world applications, the lack of labeled data will cause enormous obstacles if one hospital wants to deploy a new Sepsis detection system. Different from the supervised learning setting, we need to use known information (e.g., from another hospital with rich labeled data) to help build a model with acceptable performance, i.e., transfer learning. In this paper, we propose a semi-supervised optimal transport with self-paced ensemble framework for Sepsis early detection, called SPSSOT, to transfer knowledge from the other that has rich labeled data. In SPSSOT, we first extract the same clinical indicators from the source domain (e.g., hospital with rich labeled data) and the target domain (e.g., hospital with little labeled data), then we combine the semi-supervised domain adaptation based on optimal transport theory with self-paced under-sampling to avoid a negative transfer possibly caused by covariate shift and class imbalance. On the whole, SPSSOT is an end-to-end transfer learning method for Sepsis early detection which can automatically select suitable samples from two domains respectively according to the number of iterations and align feature space of two domains. Extensive experiments on two open clinical datasets demonstrate that comparing with other methods, our proposed SPSSOT, can significantly improve the AUC values with only 1% labeled data in the target domain in two transfer learning scenarios, MIMIC $rightarrow$ Challenge and Challenge $rightarrow$ MIMIC.