Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO

Benchmark Model Rank Results
mathematical-reasoning-on-olympiadbench4B→14B S2L#50Accuracy: 21.9
mathematical-reasoning-on-aime244B→14B S2L#109Acc: 24.4
mathematical-reasoning-on-math-5004B→14B S2L#118Accuracy: 62.7
mathematical-reasoning-on-aime-20251.7B→8B S2L#124Accuracy: 22.5