EOPD

Notes on Entropy-Aware On-Policy Distillation - an OPD method that keeps reverse-KL learning for confident teacher tokens while adding top-k forward KL at high-entropy positions to preserve reasoning diversity.