1、发现,推理长度由表示空间中的线性方向决定,从而能沿着该方向引导模型,诱导过短推理。
论文解读:ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models
未经允许不得转载:小狮博客 » 论文解读:ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models
小狮博客