Hacker News AIT2·Models
78 pts
New research introduces 'self-modeling' interventions, operationalizing a model's self-recognition and self-report capabilities to modulate Emergent Misalignment (EM). The study finds that direct measurements of these internal representations predict generalization behaviors and that interleaving self-report examples during fine-tuning effectively mitigates EM. Notably, misalignment can be transferred to fresh models solely via training on self-reports from fragmented models, offering new insights into latent mechanisms and alignment safety.