Chain of Draft (2025)
Free≤5 words per reasoning step — 91% of CoT accuracy at 7.6% of the tokens; 76% latency reduction
FreeFree tier
Inputs: textOutputs: text
About Chain of Draft (2025)
Chain of Draft (CoD) is a novel prompting paradigm for large language models that generates minimalistic intermediate reasoning outputs—limiting each step to five words or fewer. Inspired by human cognitive efficiency, CoD achieves accuracy comparable to or exceeding Chain-of-Thought (CoT) while using as little as 7.6% of the tokens, resulting in substantial reductions in cost and latency. The method is demonstrated across various reasoning tasks and is accompanied by open-source code and data.
Key Features
Generates minimalistic intermediate reasoning steps of ≤5 words each
Matches or surpasses Chain-of-Thought accuracy while using only 7.6% of the tokens
Significantly reduces cost and latency (76% latency reduction shown)
Inspired by human cognitive processes for efficient reasoning
Open-source code and data available for reproducibility
Pros & Cons
Pros
- Maintains high accuracy (91% of CoT) with drastically fewer tokens
- Reduces latency and computational cost, enabling faster responses
- Simple to implement as a prompting strategy without model fine-tuning
- Open source and well-documented for community use
Best For
Complex reasoning tasks where Chain-of-Thought is applied but cost/latency is a concernMathematical problem solvingLogical reasoning and step-by-step deductionAny NLP task requiring intermediate reasoning with efficiency constraints
FAQ
What is Chain of Draft (CoD)?
Chain of Draft is a prompting technique for large language models where the model generates concise intermediate reasoning steps (≤5 words each) instead of verbose explanations, reducing token usage while preserving accuracy.
How does CoD compare to Chain-of-Thought (CoT)?
CoD matches or surpasses CoT accuracy on various reasoning tasks while using as little as 7.6% of the tokens, leading to significant reductions in cost and latency.
Is the code for Chain of Draft publicly available?
Yes, the code and data for the paper are available at the URL provided in the arXiv submission.