SPOT: Sparse Probing and Outcome Calibration for On-Policy Distillation

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathSPOT Qwen3-4B-Base#28Accuracy: 37.73
mathematical-reasoning-on-math-500SPOT Qwen3-4B-Base#92Accuracy: 79.88
mathematical-reasoning-on-aime24SPOT Qwen3-4B-Base#123Acc: 17.5
mathematical-reasoning-on-aime-2025SPOT Qwen3-4B-Base#130Accuracy: 17.92
multiple-choice-question-answering-mcqa-on-gpqa-diamondSPOT Qwen3-1.7B-Base#131Accuracy: 29.42