Training thinking-length preferences into the model weights is more stable than using system prompts at deployment time to request concise answers, as the latter is an external constraint that is unstable while the former becomes the model's default behavior.