Recovering Reliable Supervision via Critique-Guided On-Policy Distillation

Benchmark Model Rank Results
mathematical-reasoning-on-amc23CriticOPD (Olmo-3-7B-Instruct-SFT)#3Acc: 74.7
mathematical-reasoning-on-olympiadbenchCriticOPD (Olmo-3-7B-Instruct-SFT)#18Accuracy: 62.33
mathematical-reasoning-on-minerva-mathCriticOPD (Qwen3-1.7B-Non-Thinking)#29Accuracy: 37.44
mathematical-reasoning-on-math-500CriticOPD (Olmo-3-7B-Instruct-SFT)#60Accuracy: 89.08
mathematical-reasoning-on-aime24CriticOPD (Olmo-3-7B-Instruct-SFT)#90Acc: 43.44
mathematical-reasoning-on-aime-2025CriticOPD (Olmo-3-7B-Instruct-SFT)#101Accuracy: 37.81