A new prompting technique from MIT and Stanford replaces verbose CoT with terse intermediate steps.
Chain-of-Draft (CoD), introduced by Xu et al. at MIT/Stanford, asks the model to produce terse 'drafts' of reasoning steps rather than full natural-language chains.
On GSM8K, MATH, and BIG-Bench Hard, CoD matched chain-of-thought accuracy while emitting 70% fewer tokens — a meaningful cost win for production deployments.
Chain-of-Draft (CoD) is a prompting technique from Xu et al. at MIT and Stanford. Instead of producing verbose natural-language reasoning, the model emits terse 'draft' intermediate steps that compress the same logical chain.
Across GSM8K, MATH, and BIG-Bench Hard, CoD emitted roughly 70% fewer tokens than equivalent chain-of-thought prompting, while matching CoT on final-answer accuracy.
No measurable hit on the three benchmarks reported. The authors note that the terser format encourages the model to commit to step boundaries rather than ramble, which may explain why accuracy holds.
Token costs scale linearly with output length. A 70% reduction on reasoning traces is a 70% cost cut on inference, plus correspondingly lower latency — both binding constraints at scale.