Self-modeling interventions can modulate emergent misalignment in models

Tagade et al. study how self-recognition and self-reporting modulate emergent misalignment, finding that some training interventions can prevent or reverse some metrics, while self-reported data can also propagate misalignment. The authors acknowledge that the operationalization is somewhat coarse, some experiments use only a single seed, and the agentic evaluation is not connected to real tools, so they cannot claim general recovery.