Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathFiRe-OPD#1Accuracy: 68.49
mathematical-reasoning-on-olympiadbenchFiRe-OPD#7Accuracy: 70.47
mathematical-reasoning-on-hmmt-feb-2025FiRe-OPD#21Accuracy: 35.0
mathematical-reasoning-on-math-500FiRe-OPD#32Accuracy: 94.69
code-generation-on-livecodebench-v6FiRe-OPD#60Pass@1: 28.1
mathematical-reasoning-on-aime24FiRe-OPD#60Acc: 64.17
mathematical-reasoning-on-aime-2025FiRe-OPD#78Accuracy: 55.83