Activation Probes Are Reliable With a Handful of Positive Examples
Efforts to monitor advanced AI for rare misalignments face a data challenge: abundant aligned examples but only a handful of misaligned ones.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv