Boosting Reinforcement Learning with Verifiable Rewards via Randomly Selected Few-Shot Guidance

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathFEST-GRPO#31Accuracy: 35.45
mathematical-reasoning-on-olympiadbenchFEST-GRPO#33Accuracy: 47.0
mathematical-reasoning-on-math-500FEST-GRPO#88Accuracy: 81.1
mathematical-reasoning-on-aime24FEST-DPO#118Acc: 20.41
mathematical-reasoning-on-aime-2025FEST-GRPO#137Accuracy: 14.58