Gary Marcus 质疑 Claude 数学证明的方法论
Marcus 拆解了 Claude 证明数学题的真实流程:Lean 迭代验证加人类数学家判断,不是一句 prompt 就跑出结果。想搞清 LLM 在数学证明里到底干了多少活的,可以看看这条。
Gary Marcus 在 X 上发帖,讨论一份报告提到的证明流程:将证明片段通过 Lean 定理证明器反复迭代验证,Lean 成为整个认知系统的关键部分。
他指出,数学家 Alpoge 也贡献了自身的数学判断,因此并非像部分人误以为的那样,是一个纯 LLM 接收“solve riemann”这样的提示词就直接给出证明。Marcus 还强调 Claude Code 已被证实是神经符号混合系统,他推测 Claude 本身也是如此。
这番评论针对的是近期围绕 Claude 参与数学定理证明的讨论,核心争议在于人类数学家、Lean 形式化验证与 LLM 各自在成果中占比几何,而完整方法论尚未公开。
As noted elsewhere I would really like to know what the methodology was, but e.g. one report sggests that agentic loops in which bits of proof were fed through lean and then iterated on, in which case lean was an essential part of the cognitive system that did the proof. so too was Alpoge, who also add his mathematical judgement so it’s not like a pure LLM came out with the answer from a prompt “solve riemann” in the way that some people are erroneously envisioning. of course i doubt that Claude is a pure LLM at this point; we know for a fact that Claude code is a neurosymbolic hybrid (what I long favored) and i imagine Claude is.📊 3 ⚡
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Claude, Lean, AI safety, mathematical proofs.
Source attribution: curated from Gary Marcus (@GaryMarcus).