Let the model pick its reasoning mode at the first [redacted] fewer tokens for 1.4 points of accuracy.

A 1.5B distilled model trained with GRPO chooses NoThink, Short, or Long at response start, using a shaped reward that makes each mode pay off at a different length plus hard per-mode token caps. Accuracy held at 0.782 against 0.796 baseline while mean length fell from 4,796 to 2,811 tokens, and routing transferred to easier benchmarks for 76% savings with no retraining. arXiv The per-mode token cap is the load-bearing design choice. Without it the modes collapse into one.