On policy distillation can cut training compute by up to 50 times
Unpacking the Compute Efficiency of Strong-to-Weak OPD against GRPO
Machine LearningComputation and Language
Summary
Training AI models involves teaching a smaller 'student' model by copying a bigger 'teacher' model. This copying process usually needs a lot of computing power. The authors compare two training methods and find that if the teacher is the student's own checkpoint, copying can reach the same skill but with up to 50 times less compute. They also show that using multiple teachers lets one student learn from many experts efficiently, saving lots of total training work.
What this means in practice
- •For machine learning engineers: Reduce compute costs by switching to on-policy distillation when retraining models from their own checkpoints across tasks like math and code.
- •For ai research infrastructure teams: Consolidate multiple domain expert models into a single model with much lower combined training compute using multi-teacher distillation.
Authors
Zeyu Leo Liu, Yang Li, Bo Pang, Emre Can Acikgoz, Xuan-Phi Nguyen, Anurag Koul, Srijan Bansal, Erik Nijkamp, Shafiq Joty, Semih Yavuz
Abstract
On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher's per-token supervision. Does this dense signal reduce the training FLOPs needed to reach a target accuracy compared with reinforcement learning? We systematically compare vanilla reverse-KL OPD with Group Relative Policy Optimization (GRPO) across task difficulty, student and teacher size, domain, and optimization choices, measuring held-out accuracy against training FLOPs. With an external post-trained teacher, OPD generally loses to GRPO in accuracy per training FLOP at matched settings, even when the teacher is further trained with GRPO, and often suffers response-length inflation. Yet a different teacher changes the picture: when the teacher is the student's own GRPO checkpoint, OPD recovers that checkpoint's accuracy on math with $50\times$ fewer distillation FLOPs than GRPO needs to reach the same accuracy. Following the same teacher setting, we consistently observe this high efficiency with a larger student, on code, and in a second model family. On a single domain, however, the GRPO checkpoint is already the desired model, so distilling it adds to total compute. Across domains, separate GRPO checkpoints hold different capabilities, and no single checkpoint serves every domain. Multi-teacher OPD (DeepSeek-AI, 2026) consolidates per-domain GRPO experts into a single student that matches each expert's accuracy, with the distillation stage costing only 5.2% (on 2 domains) and 8.0% (on 4 domains) of the specialist training FLOPs. Counting specialist training and distillation together, this pipeline reaches joint GRPO's performance on the 2-domain mix with 42.0% fewer total FLOPs.