KL for a KL: On-Policy Distillation with Control Variate Baseline
Open paper
Benchmark
Model
Rank
Results
multiple-choice-question-answering-mcqa-on-gpqa-diamond
Qwen3-1.7B → Qwen3-1.7B-Base with v OPD (top-k)
–
Accuracy: 28.6
Rank counts only results with a code link.