#post-training

ATOD: Hybrid Distillation for Autonomous Agent Training

ATOD combines on-policy distillation with reinforcement learning using an annealed schedule and turn-level reweighting to train small agent models that outperform their larger teacher models.