Blockwise Policy-Drift Gating for On-Policy Distillation

The paper introduces blockwise policy-drift gating, a method designed to enhance on-policy distillation (OPD) for long-horizon reasoning tasks by implementing a lightweight drift controller that operates solely on the student policy. This approach computes log-probability shifts between the behavior and current student policies over fixed blocks, improving the mean pass@8 metric from 0.4978 to 0.5160 in a six-variant Qwen3 math reasoning benchmark, suggesting that block-level gating can effectively stabilize performance in OPD scenarios. This advancement is significant for practitioners as it offers a straightforward mechanism to improve robustness in model training without altering teacher targets or rollout policies.