model reasoning

On-Policy Distillation Builds New Skills Rather Than New Knowledge

OPD improves reasoning by shaping use of stored facts, not adding knowledge; reverse-KL transfers skills, forw…