Coverage, Not Difficulty, Sets How Much Synthetic Data an Activation Probe Needs
The study investigated how much synthetic data an AI probe needs to learn different conversation types. The key finding is that the required amount of data is determined by coverage—how well samples from one conversation type cover others—rather than just the difficulty of each individual type. This means generating many kinds of samples is only necessary when no single kind covers another, and the required depth depends on what you are monitoring.