Let the model pick NoThink, Short, or Long at response start and cut mean tokens 41% for a 1.4 point accuracy loss

arXiv 2608.20256·medium signalXBlueskyLinkedInCopy link

A 1.5B distilled model was trained with GRPO to choose its reasoning mode at the first token, using a shaped reward that makes each mode pay off at a different response length plus hard per-mode token caps so the policy does not collapse into one strategy. Accuracy held at 0.782 against a 0.796 baseline while mean response length fell from 4,796 to 2,811 tokens, and the routing transferred to easier benchmarks for 76% token savings with no retraining. The transferable design lesson is the per-mode token cap; without it the modes merge.