OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification

Benchmark Model Rank Results
mathematical-reasoning-on-aime-2025OmniOPD (Qwen3-4B student, Claude-4.5-haiku teacher)–Accuracy: 65.63
mathematical-reasoning-on-aime24OmniOPD (Qwen3-4B student, Qwen3-30B-A3B-Instruct teacher)–Acc: 75.0
mathematical-reasoning-on-math-500OmniOPD (Qwen3-1.7B student, Qwen3-32B teacher)–Accuracy: 84.61
mathematical-reasoning-on-olympiadbenchOmniOPD (Qwen3-4B student, Gemini-2.5-flash teacher)–Accuracy: 64.58