Report
Self-modeling interventions may modulate emergent misalignment
A post about a new NeurIPS paper says two measures of a model’s self-model can predict the success of training interventions for emergent misalignment.
TLDR
A post introducing a NeurIPS paper says emergent misalignment has side effects on a model’s self-model. It says two simple, direct measurements of that self-model can predict the success of training interventions for emergent misalignment, and that self-modeling interventions can modulate it.
Combined views
884
2 Sources, first seen ago
