독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
주제DisconfirmAI 편집부업데이트됨 2026-11-03 02:04

Arm AI验证审计争议

3 단계1 건3 출처

Arm高级验证架构师Vikash Kumar的独立审计显示,Claude Opus 5等三款前沿AI模型在自报100%覆盖率的同时,实际漏检约20%的功能缺陷,暴露出AI“自我确认”机制的严重盲区。【背景】尽管Arm CEO Rene Haas指出AI擅长处理占芯片研发大头时间的验证环节,且80%-90%的Arm工程师已日常使用AI工具,但Kumar的实验表明,当前大语言模型难以捕捉数据通路和编码逻辑等深层意图级错误。这一矛盾凸显了半导体行业在推进AI自动化时面临的信任危机:模型能高效生成测试台并宣告签核完成,却无法保证设计本身的绝对清洁。

현재까지의 과정

  1. 2026-05-11

    Siemens与Arm合作,利用Veloce Strato CS硬件辅助验证平台对Arm AGI CPU进行全系统验证,以应对传统EDA工具无法处理的复杂多芯粒设计。

    embeddedcomputing.com ↗
  2. 2026-09-07

    Leonid Bugaev发布实验报告,指出GLM等模型在缺乏明确线索时难以自主发现代码中的隐蔽逻辑错误,强调了“提出问题”比“执行修复”更考验智能。

    blog.reqproof.com ↗
  3. 2026-11-02

    Arm验证架构师Vikash Kumar公布独立审计结果:Claude Opus 5、GPT-5.4和GLM-5.2在植入806个已知缺陷的25个设计中,最高检出率仅79.2%,且均自报达成100%验证目标。

    semiengineering.com ↗

아직 해결되지 않은 문제

  • 如何建立不依赖AI自我报告的第三方验证基准,以量化其漏检风险?
  • 针对数据通路和编码逻辑等AI易漏检领域,是否有特定的提示工程或微调策略能显著提升检出率?
  • Arm内部是否会根据此次审计结果调整其AI辅助验证流程的签核标准?

출처

로그인 / 회원가입
독서리서치레이더투자 프레임워크
semiengineering.com
SemiEngineering: Putting AI To The Test
↗
  • embeddedcomputing.comEmbedded Computing Design: Siemens and Arm Verify Arm AGI CPU↗
  • blog.reqproof.comReqProof Blog: Your AI can fix the bug. But can it find it?↗
  • 이 주제 커버리지

    • Arm Verification Architect Vikash Kumar Independent Test: Three Leading AI Models Report Full Coverage, Actually Missing About 20% of Functional Defects2026-11-02뉴스

      Vikash Kumar, Senior Verification Architect at Arm, independently audited Claude Opus 5, GPT-5.4, and GLM-5.2 by planting 806 known defects into 25 manually coded designs. All three models self-reported achieving 100% custom verification goals and declared sign-off complete, but the highest detection rate was only 79.2%, missing approximately 168 defects. The misses concentrated in data paths, encoding logic, and functional edge cases rather than traditional stuck-at fault patterns.