faster, more efficient version of CoT
For quite a few months, Chain of Thoughts has been the go-to prompting technique for any guy who is using LLM or ChatGPT. One issue that I always found with the Chain of Thoughts (CoT) was its verbosity.
https://medium.com/media/fb62ca0480d1480d3cb36953397b84b5/href
Subscribe to datasciencepocket on Gumroad
It generates just too many tokens.
You ask ChatGPT a simple question and receive an essay in return? That’s where Chain of Drafts (CoD) comes in — a prompting technique designed to streamline reasoning without sacrificing accuracy. Think of it as the “get to the point” version of the more detailed Chain of Thought (CoT) approach. Let’s break it down!
What is Chain of Drafts (CoD)?
CoD is a prompting method that prioritizes concise reasoning steps to improve efficiency. Unlike Chain of Thought (CoT), which walks through every single step in excruciating detail, CoD focuses on delivering only the essential information needed to solve a problem. This means:
- Less verbosity — No excessive explanations.
- Fewer tokens — Lower computational cost.
- Faster responses — Ideal for real-time applications.
A Simple Example
Let’s compare different prompting techniques using a basic math problem:
Problem: Jason had 20 lollipops. He gave Denny some. Now he has 12 left. How many did Jason give to Denny?
Standard Prompting (Direct Answer)
- Response: 8
Chain of Thought (CoT) Prompting (Step-by-Step Explanation)
- Response:
Jason started with 20 lollipops.
After giving some away, he had 12 left.
To find out how many he gave away, we subtract: 20–12.
That gives us 8.
Chain of Drafts (CoD) Prompting (Concise Reasoning)
- Response: 20–12 = 8 → Final Answer: 8.
See the difference? CoT is comprehensive but can be overkill. CoD trims the fat, retaining clarity while saving time.
Metrics and benchmark
As you must be able to see, CoD appears to be the real-world problem solver, being in the middle of standard and CoT prompting.
Why is CoD Better than CoT?
- Same accuracy, fewer tokens: CoD achieves nearly the same accuracy as CoT but with far fewer tokens, making it computationally cheaper and faster.
- More efficiency: Reducing token count means faster response times and lower costs in real-world applications.
- Best of both worlds: Unlike Standard Prompting, which sacrifices accuracy for brevity, and CoT, which is too verbose, CoD strikes a balance — concise yet accurate.
Advantages and Disadvantages
Advantages of CoD over CoT
Faster inference: CoD generates responses quicker due to reduced token usage.
Lower computational cost: Using fewer tokens translates to lower operational expenses.
More concise outputs: Ideal for real-time applications where brevity is key.
Easier to integrate in production systems: Less verbose reasoning makes responses more user-friendly.
Disadvantages of CoD compared to CoT
- Less transparency: CoT provides a clear, step-by-step breakdown, which is useful for debugging and explaining reasoning.
- Higher risk of errors in complex reasoning: Some problems require detailed intermediate steps to ensure logical correctness, which CoD might skip.
- Not ideal for educational purposes: When learning a new concept, detailed explanations (as in CoT) can be more beneficial.
Where CoD Works Best
- Real-time AI applications — Customer support, personal assistants, and chatbots.
- Resource-constrained environments — Running LLMs on edge devices or limited compute resources.
- Summarization tasks — Quickly distilling key points from text-heavy sources.
- Cost is priority
One last thing …
Why is it called Chain of Drafts?
The name “Chain of Drafts” (CoD) comes from the idea that instead of fully elaborating each reasoning step (like in Chain of Thought (CoT)), the model generates concise “drafts” of reasoning steps — just enough to keep the logic intact without unnecessary verbosity.
Think of it like writing a rough draft instead of a full essay. Instead of carefully spelling out every thought process in detail, CoD takes a minimalist approach, stripping reasoning down to its most essential components. This makes responses faster, more efficient, and token-economical, while still retaining accuracy.
The “chain” aspect still applies because there’s a logical progression of thoughts, just expressed in a more streamlined way.
Wrapping Up
Chain of Drafts is a simple yet powerful tweak that enhances LLM efficiency without compromising correctness. If you want responses that are faster, cheaper, and just as accurate, COD is for you.
What is Chain of Drafts? Bye Bye Chain of Thoughts was originally published in Data Science in your pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
