Translation in progress Some entries still include text in the original language.
BackSource OpenAI Alignment Research
Maja Trębacz, Sam Arnesen, Albin Cassirer, Max Johnson, Xin Lin, Thibault Sottiaux, in collaboration with the rest of the Codex team
Indexed by this site 2026-10-01 03:58 12 days ago
Originally published Dec 2, 2025
Selected 2026-10-01 03:59
Category Defenses OpenAI Alignment Research · Maja Trębacz, Sam Arnesen, Albin Cassirer, Max Johnson, Xin Lin, Thibault Sottiaux, in collaboration with the rest of the Codex team · Indexed by this site 2026-10-01 03:58 · Originally publishedDec 2, 2025 SelectedAttention50
OpenAI 分享在规模化代码验证上的实践方法 A Practical Approach to Verifying Code at Scale
AI summary
OpenAI 介绍了为 gpt-5-codex 和 gpt-5.1-codex-max 训练专用智能体代码审查器的经验,将其作为深度防御策略中的输出监控手段。审查器获得仓库级工具和执行权限后,能发现更多关键问题并减少误报,专门针对代码审查的训练进一步提升效果。部署上优先保证信噪比,宁可适度降低召回率以换取开发者信任,并允许通过自定义任务指令或 AGENTS.md 调整这一取舍。数据显示,Codex 云完全生成的 PR 中有 36% 收到评论,其中 46% 促使作者修改代码,人工编写 PR 的评论修改比例为 53%;截至 2025 年 10 月,系统每天处理超过 10 万条外部 PR,超过 80% 的评论反馈为正面。
Why it was selected
OpenAI 公开了代码审查智能体的训练取舍与部署数据,可了解输出监控在真实工程流程中的落地方式。
Read original alignment.openai.com Why it was selected SelectedAttention50
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
OpenAI 公开了代码审查智能体的训练取舍与部署数据,可了解输出监控在真实工程流程中的落地方式。