研究：激活探针所需合成数据量由覆盖度而非难度决定

1. 12:00
   arXiv · Machine Learning Theory

   覆盖度而非难度决定激活探针所需合成数据量

   激活探针在部署语言模型监控中依赖合成对话训练，所需样本量由监控概念决定。研究追踪 10-590 个合成样本的学习曲线，在 14 个留出评测分布和 4 个探针模型上发现，Gemma-3-27B-IT 上高风险与有害回复探针在 80 样本即接近平台期，指令遵循探针则需数倍样本。

Current heat 9·Comparable peak 10（Oct 9）·Comparable change over 24 hours –

The trend compares only the same participants observed continuously; its range may be smaller than the current heat count. Move or click on the chart to inspect hourly heat; use the left and right arrow keys to switch.