arXiv · Computation and Language· Wenxi Gan·· 4 hr agoAI score29
Learning to Act with Task Progress:基于紧凑教师监督蒸馏小型智能体
Learning to Act with Task Progress: Distilling Small Agents from Compact Teacher Supervision
AI brief
论文提出 Task-Progress Distillation(TPD),一种离线方法,为教师轨迹中的每个动作配上描述当前任务阶段的简短标签,学生模型通过联合评分合法阶段—动作对来选择动作。
Source: arXiv · Computation and Language · arxiv.org