Large language models that reason step by step are accurate but slow and expensive, even for simple questions. Systems that answer instantly save time but may be less accurate. This work asks whether one model can do both, choosing the right mode per question.
The authors create a hybrid model with a frozen base and a small routing component. The router estimates how much value a full reasoning pass would add over a quick answer. If that value exceeds a per-request cost threshold, the model runs reasoning; otherwise it answers without generating any tokens.
On several tests, the hybrid model matched or slightly beat existing systems while cutting inference time by about three-quarters and reducing token use by a similar margin. This shows that models can be both fast and accurate by learning when to think.