推荐leetrout · 2025-12
“Related check out chain of draft if you haven't.”
推荐一看 chain of draft,用简短草稿达到相近效果,只用 7% 的 CoT token
让 LLM 只写简洁草稿式推理,token 用量大减
针对思维链(CoT)提示中冗长的逐步推理,论文提出受人类认知启发的草稿链(CoD):让 LLM 在解题时只产生极简而有信息量的中间推理。在多种推理任务上,CoD 的准确率达到或超过 CoT,而 token 使用量最低只有 CoT 的 7.6%,显著降低成本和延迟。
评论者把它当作「用少得多的 token 获得相近推理效果」的相关工作来推荐,同时有人指出另一项工作与它几乎相同,但做了训练而不只是靠提示。
“Related check out chain of draft if you haven't.”
“This seems near identical to chain of draft, albeit more thorough and they actually did training instead of just promoting Chain of Draft”
“Similar level of results in a fraction of the tokens resulting in similar quality for less cost for longer runs.”