Determine the eventual scaling limit of white-box probes

Determine whether scaling white-box probes with additional training data eventually reaches a limit in performance, thereby characterizing the long-run benefits of increasing probe-training data.

Background

The paper demonstrates that increasing both the quantity and diversity of the FIBS training dataset improves the performance of white-box probes for detecting falsehoods and monitoring agentic sabotage. The probes also generalize from contextual falsehoods to introspective deception and perform increasingly well as the underlying model is scaled.

Despite these results, the reported experiments do not establish whether the observed gains from adding training data will continue indefinitely or eventually saturate. The authors explicitly identify this as a future-work question concerning the existence of an eventual performance limit for data scaling in white-box probes.