CMU: Merging Plus Grafting Lets Post-Trained Models Absorb Off-Policy Data Without Forgetting

CMU researchers introduce a "grafting" method based on model merging that lets post-trained models absorb off-policy data without forgetting old capabilities, challenging the view that continual learning must be on-policy. It outperforms self-distillation baselines in experiments.

2026-11-03 ~ 2026-11-03 · 2 related posts

• CMU: Model Merging and 'Grafting' Let Post-Trained LLMs Learn Off-Policy Without Forgetting — AdtRaghunathan · 2026-11-03
• Grafting: model merging after off-policy SFT beats on-policy self-distillation for continual learning — AdtRaghunathan · 2026-11-03