독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
주제DisconfirmAI 편집부업데이트됨 2026-11-07 10:25

Lean形式化验证信任

6 단계2 건7 출처

【背景】Lean定理证明器创始人Leonardo de Moura近期披露,两个独立检查器曾分别被不同漏洞欺骗,揭示了小可信内核架构下的潜在风险。与此同时,随着AlphaProof等AI系统生成无人能读的证明,“形式化差距”成为焦点:机器验证了逻辑步骤,却未必验证了人类意图。

最新研究显示,AI生成的Lean代码存在显著的“编译-忠实性”鸿沟。例如,某增强型GPT-5.2代理虽使87.2%的陈述成功编译,但仅57.7%真正符合原始数学意图。这意味着“代码通过编译”不等于“证明了用户想证的定理”。

【存疑】尽管AWS等机构利用Lean验证Cedar授权策略的安全性,且社区开发了lean4checker等工具以防范恶意构造,但核心争议在于:当AI同时编写规格说明与证明时,如何确保规格说明本身未因误解而偏离真实需求?目前尚无定论,需依赖人工审查与独立校验器的交叉验证。

현재까지의 과정

  1. 2013-01-01

    Leonardo de Moura在微软研究院启动Lean项目,最初旨在验证软件代码正确性。

    sciencenews.org ↗
  2. 2024-07-01

    Google DeepMind的AlphaProof在国际数学奥林匹克竞赛中作为非官方参赛者获得银牌水平成绩,引发对AI自动推理能力的关注。

    sciencenews.org ↗
  3. 2026-09-30

    Leonardo de Moura在播客中讲述Collatz证明事件中,两个独立检查器分别被不同漏洞欺骗的过程,强调小可信内核与独立校验器的重要性。

    podcasters.spotify.com ↗
  4. 2026-10-02

    加州大学河滨分校等机构发布预印本《Beyond Compilation》,量化AI自动形式化的“编译-忠实性”差距,指出部分模型编译率高但语义忠实度低。

    arxiv.org ↗
  5. 2026-10-23

    Cairn Commons发布指南,详细列出检查AI生成Lean证明的五步法,包括核对正式陈述、排查`sorry`、打印公理、本地构建及对比预设陈述。

    cairn-commons.com ↗
로그인 / 회원가입
독서리서치레이더투자 프레임워크
  • 2026-11-07

    Tony Bai撰文分析Leo de Moura观点,指出AI正推动形式化验证进入“ImageNet时刻”,但强调模型忠实性(Faithfulness)仍是关键瓶颈。

    tonybai.com ↗
  • 아직 해결되지 않은 문제

    • 如何自动化检测AI生成的Lean陈述是否准确反映了自然语言描述的数学意图,而非仅仅语法正确?
    • 在AI同时生成规格说明和证明的场景下,是否存在不依赖人工介入的独立第三方机制来防止“自我实现”式的错误验证?
    • Lean内核漏洞#14576等历史案例表明,现有外部检查器(如lean4checker)能否完全覆盖所有类型的恶意或意外构造?

    출처

    • sciencenews.orgScience News: The Proof in the Code 书评↗
    • podcasters.spotify.comMLST Podcast: Leo de Moura 访谈↗
    • arxiv.orgArXiv: Beyond Compilation (AI形式化忠实性研究)↗
    • cairn-commons.comCairn Commons: 如何检查AI Lean证明↗
    • lean-lang.orgLean Lang: 验证证明参考文档↗
    • aws.amazon.comAWS Blog: Lean用于Cedar验证↗
    • tonybai.comTony Bai: AI写代码谁证明它有效↗

    이 주제 커버리지

    • Kevin Hartnett, Author of 'The Proof in the Code': How the Formalisation Gap Explains AI Math Verification Challenges2026-10-27뉴스

      Machine Learning Street Talk releases a documentary centered on Kevin Hartnett's new book, 'The Proof in the Code'. The film explores the evolution of the Lean theorem prover from a Microsoft Research side project to a core tool for mathematicians, and delves into the 'formalisation gap' raised by AI systems like AlphaProof: when a machine produces a proof nobody can read, what has actually been checked?

    • Lean Creator Leonardo de Moura: How a Claimed Proof Suspectedly Exploited Different Vulnerabilities to Bypass Two Independent Kernels2026-10-26뉴스

      In the September 30, 2026 episode of Machine Learning Street Talk, Lean theorem prover creator Leonardo de Moura recounts how two independent checkers were deceived by different vulnerabilities in the Collatz proof incident, and discusses the architecture design of small trusted kernels and independent verifiers.