Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs
Draft-Thinking improves large reasoning models by training them to learn a concise 'draft' reasoning style that captures only critical steps. This is achieved through supervised fine-tuning followed by progressive reinforcement learning, which stabilizes this efficient pattern. The method allows models to choose between fast draft mode and detailed long CoT mode based on task difficulty, significantly reducing inference cost without sacrificing accuracy.