让变化经过筛选，
再抵达你。

研究者提出一种受量子概率启发的信任估计框架，并把模型输出与 15 名参与者给出的 349 次自评进行比较；报告称两者走势接近，0–100 量表上的平均偏差为 −0.78。论文将其定位为概念验证。样本很小，自评也不等同于客观安全行为，因此不能据此认定系统已能可靠读取用户信任。

Researchers propose a quantum-probability-inspired trust model and compare its estimates with 349 self-reports from 15 participants. They report close tracking and a mean bias of −0.78 on a 0–100 scale. The paper presents a proof of concept: the sample is small, and self-reported trust is not objective safety behavior, so the result does not establish reliable trust reading in real deployments.

一项面向中国一所高职院校 884 名学生的问卷研究使用结构方程模型分析英语课堂中的 AI 能力、态度、使用经验、创意任务与学习互动。论文报告 AI 融入和创意任务参与直接预测互动，其余因素通过二者产生间接关系。研究是单校、自陈和观察性设计，只能说明变量关联，不能证明 AI 导致学习提升。

A questionnaire study of 884 students at one Chinese higher vocational college uses structural equation modeling to examine AI competency, attitudes, experience, creative tasks, and interaction in English classes. It reports direct predictive relationships for AI integration and creative-task involvement, with indirect paths for the other factors. The single-site, self-reported, observational design shows associations, not that AI caused better learning.

研究使用非接触式单通道心搏冲击图信号、心率与呼吸变异特征，以及双层分类模型，在 15 份独立记录的 16,660 个片段上完成五类睡眠分期；作者报告准确率 82.9%。这是接受后提前公开的同行评议版本，但独立记录数量有限，尚不能替代多导睡眠监测或临床诊断。

The study combines non-contact single-channel ballistocardiogram signals, heart and respiratory variability features, and a two-layer classifier for five-stage sleep classification. Across 16,660 segments from 15 independent recordings, the authors report 82.9% accuracy. This is an accepted peer-reviewed early version, but the small number of independent recordings means it cannot replace polysomnography or clinical diagnosis.

研究在 NIH CXR14 数据集上测试经粒子群优化的 MobileNetV2–ViT 混合架构，对 14 类胸部疾病做多标签分类；五折交叉验证报告平均准确率 94.6%、F1 为 91.1%，并用 Grad-CAM 展示关注区域。这是数据集内实验结果，不是前瞻性临床验证，也不应直接理解为可用于诊断。

Researchers test a particle-swarm-tuned MobileNetV2–ViT hybrid on the NIH CXR14 dataset for 14-label chest-disease classification. Five-fold cross-validation reports 94.6% mean accuracy and a 91.1% F1 score, with Grad-CAM visualizations. These are within-dataset experimental results, not prospective clinical validation or evidence that the model is ready for diagnosis.

研究者在 UK Biobank 的 16 项表型上，用梯度提升树先从时间、地点和其他非遗传协变量预测表型，再把该预测加入多基因评分流程。论文报告全部表型均改善，BASIL 方法的 R² 中位提升 7.3%。结果来自特定生物样本库与建模流程，迁移到其他人群前仍需外部验证。

Using 16 UK Biobank phenotypes, researchers first predict traits from time, location, and other nongenetic covariates with gradient-boosted trees, then add that prediction to polygenic-score workflows. The paper reports improvements for all phenotypes and a median 7.3% R² gain with BASIL. Because the finding comes from one biobank and modeling setup, external validation is needed before transfer to other populations.