loss-smoothing

Loss smoothing interpolates between source and target objectives during adaptation, preserving useful features while enabling specialization. Experiments across supervised shifts, RL, and language model fine-tuning show consistent improvements.