Coverage, Not Difficulty, Sets How Much Synthetic Data an Activation Probe Needs

Title: Coverage, Not Difficulty, Sets How Much Synthetic Data an Activation Probe Needs

标题：覆盖率而非难度决定了激活探针所需合成数据的数量

Activation probes that monitor deployed language models are trained on synthetic conversations, and how many a probe needs is open. We trace learning curves over 10-590 synthetic samples for three monitoring concepts, high-stakes situations, replies harmful to a person, and replies that do not follow the user’s instruction, on fourteen held-out evaluation distributions and four probe models, varying the generator LLM and the prompt’s detail.

用于监控已部署语言模型的激活探针通常在合成对话上进行训练，但探针具体需要多少数据尚无定论。我们针对三个监控概念（高风险情境、对个人有害的回复，以及未遵循用户指令的回复），在十四个留出的评估分布和四个探针模型上追踪了 10 到 590 个合成样本的学习曲线，并改变了生成器大语言模型（LLM）和提示词的细节。

The need is set by what is monitored: probes for high-stakes and harmful are within a few hundredths of their plateau from 80 samples on Gemma-3-27B-IT, instruction probes need several times as many, and the ordering holds on three smaller probe models and on real samples (from dev set). Prior work advises spending a generation budget on breadth, more kinds of data, over depth, more of each kind.

需求取决于监控的对象：在 Gemma-3-27B-IT 模型上，针对高风险和有害内容的探针在 80 个样本时已达到其平台期的百分之几以内，而指令探针则需要多出几倍的数据量；这种排序在三个较小的探针模型和真实样本（来自开发集）上同样成立。先前的研究建议将生成预算投入到广度（更多种类的数据）上，而非深度（每种数据更多数量）。

We read the depth a concept needs as the half-gain size of a fitted curve, the number of samples at which half the gain is in hand. Concept and distribution account for 42-45% of its variance, the generator, probe model, and prompt detail for under 10%. What sets the value of the half-gain size is coverage, not per-kind difficulty: the number of samples of its own kind a distribution needs to saturate.

我们将一个概念所需的深度解读为拟合曲线的“半增益大小”，即获得一半增益时所需的样本数量。概念和分布占其方差的 42-45%，而生成器、探针模型和提示词细节的影响不到 10%。决定半增益大小的是覆盖率，而非单类难度，即一个分布在饱和时所需的同类样本数量。

Every kind, one per evaluation distribution, has a median half-gain size of 7-11 own-kind synthetic samples under all three concepts alike. What differs is how far samples of one kind transfer to the concept’s other kinds, almost fully under high-stakes, less under harmful, and least under instruction, which accounts for most of the gap between concepts on generated and real samples.

每种类型（每个评估分布对应一种）在所有三个概念下，其中位半增益大小均为 7-11 个同类合成样本。不同之处在于，一种类型的样本在多大程度上能迁移到该概念的其他类型中：在高风险概念下几乎完全迁移，在有害概念下迁移较少，而在指令概念下迁移最少，这解释了生成样本和真实样本在不同概念间的大部分差距。

Breadth therefore pays differently by concept: many kinds are necessary under instruction, where no kind covers another, and nearly redundant under high-stakes, where one kind covers the rest. We release the evaluation suites, dev sets, and generated sets.

因此，广度的价值因概念而异：在指令概念下，由于没有任何一种类型能覆盖其他类型，因此多种类型是必要的；而在高风险概念下，由于一种类型就能覆盖其余类型，广度几乎是多余的。我们发布了评估套件、开发集和生成数据集。