Weak-to-Strong On-Policy Distillation

Benchmark Model Rank Results
mathematical-reasoning-on-hmmt-feb-2025W2S-OPD (single-teacher contrast)#20Accuracy: 35.1
instruction-following-on-ifbenchW2S-OPD (pre-/post-RL contrast)#46Accuracy: 27.0
mathematical-reasoning-on-aime24W2S-OPD (single-teacher contrast)#50Acc: 68.9
code-generation-on-livecodebench-v6W2S-OPD (multi-teacher contrast)#61Pass@1: 27.1
mathematical-reasoning-on-aime-2025W2S-OPD (multi-teacher contrast)#68Accuracy: 61.0
multiple-choice-question-answering-mcqa-on-gpqa-diamondW2S-OPD (pre-/post-RL contrast)#104Accuracy: 56.5