Does On-Policy Distillation for Safety Pose Backdoor Risks?
On-policy distillation (OPD) has attracted growing attention as an effective way to transfer capabilities from teacher models to student models. Recent studies further explore OPD as a tool for improving large language model safety with promising results. However, these approaches typically assume that the teacher and...
Jian Luo, Ke-Han Qi, Qing-Qiao Hu et al.
· 0 citations