Tag
This paper proposes compressing reasoning traces before knowledge distillation to reduce computational costs and inference lengths, showing an accuracy-efficiency trade-off where compressed traces retain up to 96% of raw-trace accuracy with up to 18x higher per-token efficiency.