摘要
针对大语言模型生成内容难以区分的问题，现有检测器往往缺乏透明度且在分布偏移下表现不佳。本文提出 READER，一种推理增强的 AI 文本检测器，能在输出判定前先生成结构化的推理依据。通过构建包含推理过程和结论的监督数据集 READ，并对 15 亿参数模型进行微调，READER 在推理时实现了“先推理后检测”。实验表明，其性能持续优于现有检测器及规模大数百倍的提示式大模型基线。
AI 推荐理由
论文核心提出基于推理增强的检测器，利用思维链机制提升性能。
研究机构
Department of Statistics, London School of Economics and Political Science, Email: p.su@lse.ac.uk
Department of Statistics, London School of Economics and Political Science, Email: K.Ye@lse.ac.uk
School of Management, University of Science and Technology of China, Email: shijin49@mail.ustc.edu.cn
School of Mathematics, University of Birmingham, Email: j.zhu.7@bham.ac.uk
Department of Statistics, London School of Economics and Political Science, Email: j.zhu.7@bham.ac.uk
Department of Statistics, London School of Economics and Political Science, Email: G.Livieri@lse.ac.uk
Department of Statistics, London School of Economics and Political Science, Email: C.Shi@lse.ac.uk
论文信息