Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training
2026-05-12 • Machine Learning
Machine LearningArtificial Intelligence
AI summaryⓘ
The authors explain that when training AI models with limited verified data, it's better to first use complex models to learn from sparse, high-level rewards and then distill that knowledge into smaller models using detailed, dense feedback. They tested this approach with math problem-solving models and found that teaching a smaller model through a 'bridge' method—transferring knowledge from a bigger, smarter model—works better than training the small model directly with sparse rewards. Their method helps the smaller models learn more effectively and improves performance. Overall, they suggest using sparse signals for discovery on big models and dense feedback for compressing knowledge into smaller ones.
sparse rewardsdense supervisionmodel distillationreinforcement learningGRPOOPDreward densityteacher-student modelknowledge transferMATH dataset
Authors
Yuanda Xu, Hejian Sang, Zhengze Zhou, Ran He, Zhipeng Wang, Alborz Geramifard
Abstract
In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated carefully. The standard practice is to use this data directly on the model that will be deployed, for example by running GRPO on the deployment student. We argue that this is often an inefficient allocation because it overlooks a reward-density principle: sparse sequence-level reward should train models where exploration is productive, while dense token-level teacher reward should be used where the aim is to compress behavior into a smaller model. In this view, GRPO-style sparse RL and OPD-style dense teacher supervision are not separate recipes; they are different reward-density regimes. The allocation rule is simple: use scarce labeled training data upstream on the strongest model that can turn it into reward-shaped behavior, then transfer that behavior downstream as dense supervision. We evaluate this rule on verifiable math with Qwen3 and Llama models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student, while transfer from the same teacher before RL underperforms. The bridge is important: a forward-KL warmup on teacher rollouts followed by OPD on student rollouts is consistently strongest on MATH before any post-bridge student-side sparse RL, and also gives the best pre-Stage~3 AIME endpoints for the canonical 8B/14B teachers. The bridge also makes later student-side sparse RL effective: GRPO that is weak on a cold student lifts MATH from $75.4\%$ to $78.5\%$ after the bridge and outperforms a matched replay control by $2.8$ points. The operational principal is to avoid using scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge.