引用jarrattp31 · 2025-11
“Our new research demonstrates the first system to complete 1M+ dependent reasoning steps with zero errors”
介绍这项新研究首次零错误完成百万步以上相互依赖的推理
MAKER:极端拆分任务加多智能体投票,百万步零错误
大模型有持续的错误率,在汉诺塔等任务上最多几百步就会出错脱轨。论文提出 MAKER,称是首个成功完成超过一百万个 LLM 步骤且零错误的系统,原则上还能扩展得更远。做法是把任务极端分解为许多子任务,交给专注的微智能体处理,高度模块化使得每一步都可以用高效的多智能体投票方案做纠错。作者据此认为,与其依赖 LLM 的持续改进,不如采用大规模分解的智能体流程(MDAP)来解决组织与社会层面的问题。
HN 评论者引它作为“通过任务分解让 LLM 完成百万步依赖推理且零错误”的研究,用来说明如何拆分任务并合并多个模型的答案。
“Our new research demonstrates the first system to complete 1M+ dependent reasoning steps with zero errors”
“Its based on a paper* that was featured on here a while back actually [1].”
“There is research[0] currently being done on how to divide tasks and combine the answers to LLMs.”