Concise Reasoning Can Match Verbose Reasoning
Chain of Draft (CoD) prompting generates minimal, information-dense outputs at each reasoning step, and matches or exceeds Chain-of-Thought performance while using far fewer tokens.
The insight mirrors human problem-solving: when working through a math problem, we jot down key values and operations, not full sentences explaining each step.
Results:
- Comparable accuracy to CoT across arithmetic, commonsense, and symbolic reasoning
- 80% reduction in output tokens in some cases
- 76.2% average latency reduction
This challenges the assumption that more explicit reasoning is always better. What matters is capturing the information needed at each step, not the verbosity of expression.
Practical implication: for production systems where compute cost and latency matter, verbose reasoning may be unnecessary overhead.