Tag
The paper introduces Influence-Directed Adaptive On-Policy Distillation (IDA-OPD) to solve the diversity bottleneck in sampled-token on-policy distillation, enhancing diversity transfer in reasoning models without full-vocabulary teacher data.