D²-OPD on-policy distillation

Tag "D²-OPD on-policy distillation" — all articles

062026-11Two arXiv Papers Same Day: Decoupling On-Policy Distillation and Training-Free AuditingOn Nov 6, arXiv's cs.LG section published two large-model studies. D²-OPD proposes an on-policy distillation method that decouples direction from distance, beating the strongest baseline by up to 9.45 points on math reasoning benchmarks; CMaG reveals that recognizing and refusing compositional harm in vision-language models are decoupled, and proposes a training-free auditing method that reduces compositional harm on Qwen2.5-VL-7B and LLaVA-1.5-7B.arXiv cs.LG