Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO
Researchers have introduced a new method called S2L-PO that uses smaller language models to improve the training of larger ones. This approach leverages the inherent policy-level diversity of smaller models, which leads to more coherent and structured exploration during training compared to simply adding token-level randomness. By using smaller models as natural explorers, S2L-PO can enhance performance on benchmarks like mathematical reasoning while also reducing the computational cost of training. AI
IMPACT Introduces a novel training paradigm that enhances LLM performance and efficiency through diverse exploration.