The research group Resolution seeks to identify and control low-dimensional structures that emerge during pretraining. This approach targets phenomena like subliminal learning and emergent misalignment. By isolating these personas, researchers aim to intervene on model behavior without triggering hidden undesirable traits. This technical focus provides a specific path for preventing superintelligent agent misalignment.