arXivAI Research
Strategically Diverse Sampling: Why Approach Diversity Beats Model Scale in Self-Training
自我訓練新突破：策略多樣性取樣，如何讓小模型靠「多元思路」擊敗大模型蒸餾？
This study introduces "strategic diversity" sampling (GROOT and VS), proving that diverse reasoning approaches in self-training are more effective than relying on correctness or distillation from massive teacher models.
2 min read