EVENT DOSSIER
On September 7, 2026, arXiv cs.CL published the paper “Labels have Human Values: Value Calibration of Subjective Tasks”. This study addresses the lack of objective standards in subjective tasks and proposes a value calibration method. The method aims to incorporate human values into the labels generated by the model, thereby addressing the challenge of traditional evaluation systems’ inability to quantify subjective qualities such as artistic merit and emotional resonance. By introducing quantifiable value metrics, the study seeks to ensure that the output results comply with human ethical and aesthetic standards while maintaining the flexibility of the model, providing a new approach for enhancing the reliability of large models in creative tasks.
提出 MultiCalibrated Subjective Task Learning (MC-STL) 框架以解决 NLP 模型在主观任务中忽略潜在价值结构导致预测校准偏差的问题。该框架通过标签理由相似性、专家价值分类或标注者社会文化描述符识别潜在价值组，并实施基于价值组的条件校准。MC-STL 适用于二元、序数和偏好学习设置，并在有毒聊天机器人对话、价值推理及 T2I 安全与偏好对齐等多个数据集上进行评估。实验结果表明，MC-STL 在相关价值组中实现多校准，同时提升了概率预测性能，表现优于现有基线模型。