What happens to a model’s persona over many generations of self-distillation?
This study tracks the change with latent measurements, behavioural evaluations and causal interventions, and tests whether anchoring to the original model with a KL term slows the drift without costing general ability.