Move Over Chain of Thought | The Rise of Chain of Draft in AI Reasoning
The AI world is always evolving, and one of the biggest game changers in recent years has been Chain of Thought (CoT) prompting. This structured reasoning approach has propelled AI models to new heights by allowing them to break down complex problems step by step. However, while CoT has proven effective, it comes with significant drawbacks mainly, its verbosity and high computational costs. But now, there’s a more efficient alternative: Chain of Draft (CoD).
The Problem with Chain of Thought
CoT enables AI models to think step by step, mimicking human structured reasoning. This has led to breakthroughs in reasoning based tasks, but it also requires models to generate an extensive number of tokens. The result? Higher latency and increased computational costs.
Thinking models like DeepSeek 2, Gemini, and others that rely on CoT often scale their computations significantly at inference time. While CoT provides thorough explanations, it doesn’t necessarily reflect how humans approach problem solving. In reality, we often rely on concise drafts or shorthand notes to capture essential insights without unnecessary elaboration.
Introducing Chain of Draft: A More Efficient Approach
Researchers at Zoom Communications (yes, the video conferencing company) have proposed a new strategy Chain of Draft. This novel prompting technique achieves results comparable to or even better than CoT while being significantly more efficient in both cost and speed.
Rather than generating verbose intermediate steps like CoT, Chain of Draft encourages language models to create concise, dense information outputs at each step. This allows the model to maintain transparency in its reasoning process without the excessive overhead of traditional CoT reasoning.
How Chain of Draft Works
To illustrate, consider this simple math problem: Question: Kumar had 20 lollipops. He gave Namal some lollipops. Now Kumarhas 12 lollipops. How many lollipops did Kumar give to Namal?
Standard Model Response:
Chain of Thought Response:
While CoT ensures transparency, it includes a lot of unnecessary steps for a simple problem.
Chain of Draft Response:
With CoD, we get the essential reasoning needed without excessive elaboration, reducing token usage and processing time.
Performance Comparison: CoT vs. CoD
The researchers tested CoD on multiple benchmarks and compared it to standard prompting and CoT.
Implementing Chain of Draft: Simpler Than You Think
One of the most exciting aspects of CoD is that it doesn’t require any model fine tuning, reinforcement learning, or architectural changes. It is purely a prompting strategy, meaning it can be implemented immediately by modifying the system message.
Example Prompts:
The Future of AI Reasoning
The introduction of Chain of Draft shows that small changes in how we guide AI models can lead to massive efficiency improvements. While CoT revolutionized AI reasoning, CoD refines it further delivering nearly the same accuracy at a fraction of the cost and time.
In a world where AI models are increasingly integrated into real time applications, reducing latency and computational expenses is crucial. Chain of Draft provides a powerful yet simple solution to this problem, proving that sometimes, less is more.
What do you think about Chain of Draft? Could this be the future of AI reasoning?