Tag
Maxime Labonne shares Aayush Karan's research arguing SFT generalizes worse than RL because of off-policy data rather than the objective itself — rewriting expert trajectories to match the base model's own outputs can rival RL and on-policy distillation while forgetting less.