Ever felt like AI models could be a bit verbose? You’re not alone. There’s a new approach shaking things up in the AI world, and it’s all about keeping things short and sweet.
It’s called “Chain of Draft” (CoD), and it’s transforming how AI handles complex tasks by cutting down on unnecessary words without losing accuracy. Imagine summarizing your thoughts in just a few words—CoD does just that for AI, making it much more efficient.
Research shows that CoD can tackle problems using up to 92.4% fewer words than the traditional Chain of Thought (CoT) method. How does it pull this off? By limiting each reasoning step to just five words.
In tests involving arithmetic, comprehension, and symbolic reasoning, CoD didn’t just match the accuracy of detailed CoT prompts—it slashed word usage by 68 to 86 percent. Impressive, right?
Take date comprehension tasks, for example. CoD boosted accuracy from 72.6% to 88.1% for GPT-4o and from 84.3% to 89.7% for Claude 3.5 Sonnet.
By generating shorter reasoning steps, CoD reduces both the number of output tokens and input token counts, especially in few-shot prompting scenarios. This means lower computational costs and faster response times, making CoD a great choice for large-scale AI applications where cost-efficiency is a big deal.
But let’s not get ahead of ourselves. Compact prompts aren’t perfect. Some tasks need more in-depth consideration, self-correction, or even external information. To tackle these challenges, researchers suggest combining CoD with adaptive parallel reasoning or multi-level validation techniques.
These insights could even shape the way future AI models are trained by incorporating compact reasoning strategies into training datasets.
Developed by the research team at Zoom Communications, known for their “AI Companion” meeting assistant launched in 2023, CoD is set to enhance real-time AI applications, like video conferencing, by significantly reducing response latency.
