The Chain of Draft (CoD) method is an innovative prompting technique that helps language models generate clear, well-reasoned responses while minimizing token usage. It was developed by the research team at Zoom Communications. Instead of producing a final answer in one step, CoD guides the model through multiple iterative drafts, refining the response until it is both precise and cost-effective. In this approach, it borrows from how humans tackle problems.

How It Works

1. Initial Draft: The model first generates a rough version of the answer, capturing key ideas without concern for brevity, e.g. 'with maximum 5 words'.

2. Iterative Refinement: The model then reviews and condenses its response, eliminating unnecessary details and focusing on clarity.

3. Final Answer: The process continues until the response is concise, optimized, and easy to understand.

Benefits of the CoD Method

- Improved Reasoning: By working through drafts, the model engages in step-by-step thinking without unnecessary verbosity.

- Efficient Token Usage: Reducing wordiness helps lower computational costs while maintaining response quality.

- Better Output Control: Users can fine-tune the depth and length of responses by adjusting the number of refinement steps.

Why It Matters

CoD is particularly useful in applications where clarity and efficiency are crucial, such as summarization, research, and business communications. By balancing depth and conciseness, this method enhances the effectiveness of AI-generated content without unnecessary computational overhead.

For businesses and professionals leveraging AI, the Chain of Draft method provides a structured way to generate high-quality insights quickly and affordably.

To learn more, check out the paper Chain of Draft: Thinking Faster by Writing Less by Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He.