How do you correctly pass a large RL-trained teacher's "change" on to a smaller student? The answer turned out to be hiding in the training process, not just the endpoints. Title: Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation URL: Researchers at Alibaba Cloud propose Waypoint-OPD, a distillation method that uses an intermediate checkpoint from the teacher's RL training to sharpen what gets transferred to the student. Highlight 1 🧭 Checking direction, not just the endpoint Standard distillation only compares before and after training. Waypoint-OPD adds a mid-training checkpoint and measures, via cosine similarity, whether the first-half and second-half policy shifts point the same direction. Tokens where the directions disagree get automatically down-weighted. Highlight 2 📈 Consistent gains across math and code Across Qwen3-8B/14B/30B-A3B, it beats Direct-OPD on all 7 math benchmarks and all 5 code benchmarks tested, improving by 1.59-2.18 points on math and 1.19-2.66 points on code. Qwen3-14B even surpasses its own teacher model on all 7 math benchmarks. Highlight 3 ⚙️ Lightweight and robust The only extra cost is one additional scoring pass through the frozen teacher, no new trainable parameters or verifier calls needed. Moving the mid-training checkpoint between RL steps 500, 600, and 700 changes the score by only 0.05 points, so the exact choice barely matters. Exploiting the direction of training dynamics, something endpoint-only comparisons simply can't see, improves distillation quality at almost no extra compute cost. #Distillation# #ReinforcementLearning#