Preprint
Aug 2026
When Teachers Mislead: Spurious-Signal-Aware On-Policy Distillation
SA-OPD is proposed, a Spurious-Signal-Aware On-Policy Distillation framework that identifies and filters misleading token-level supervision based on input-groundedness and optimization impact, and consistently outperforms Vanilla OPD and competitive selective methods.
Yinuo Jiang, Yongjie Ye, Zhou Tao et al.
· 4 citations