Extremely Sparse Supervision Incentivizes Reasoning Ability

Based on the traditional assumption that post-training large language model training relies on massive amounts of dense supervised tokens, the research team found in the on-policy distillation (OPD) setup based on the Qwen3 family that a very small proportion of generated tokens (only one or two per inference trajectory, accounting for 0.05% of the total) can effectively stimulate the model’s reasoning ability. Experiments show that this sparse supervision approach often matches or even surpasses the performance of full-token training in most cases. This phenomenon was observed in teacher - student configurations with nine different model sizes and was further confirmed in mathematical reasoning tasks, code-based reasoning, and reinforcement learning based on verifiable rewards (RLVR). The research results challenge the traditional view that post-training must rely on dense tokens, suggesting that extremely sparse supervision may be closer to the natural learning process.