OPSDC: On-Policy Self-Distillation for Reasoning Compression

arXiv 2603.05433·medium signal

Teaches reasoning models to be concise by self-distilling: condition the same model on "be concise" to get teacher logits, minimize reverse KL on student rollouts. No ground-truth answers or token budgets needed. On MATH-500 with Qwen3-8B/14B: 57-59% token reduction with 9-16 point accuracy improvement. On AIME 2024: 10 point gain with 41% compression. Directly addresses the "reasoning tax" problem — most CoT tokens are noise that compounds errors. Relevant to Anthropic's distillation defense th