Alibaba Cloud's Waypoint-OPD Gates RL Distillation Using Mid-Run Teacher Checkpoints, Tops Six Qwen3 Settings

Alibaba Cloud researchers introduce Waypoint-OPD, a distillation method that compares a teacher model's early- and late-RL policy shifts at a mid-run checkpoint and gates endpoint supervision when the two shifts disagree in direction. Using RL-trained Qwen3-4B teachers, Waypoint-OPD posts the highest domain-average score across all six math and code reasoning settings with Qwen3-8B, Qwen3-14B and Qwen3-30B-A3B students, needing only one extra frozen-teacher scoring pass. Ablations show removing or shuffling gates degrades math reasoning.