SWE-Bench Pro V2：更干净、更难作弊的编码基准
原标题：SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark
AI 摘要
Scale AI Research 发布 SWE-Bench Pro V2，对原 731 个公开任务进行审计与修复，最终保留 642 个任务并新增 51 个任务的 Hard 子集。23 名工程师投入 404 小时修复或淘汰 480 个任务，并通过限制网络访问、在干净容器中重新评分来防止作弊。结果显示 Claude Opus 5 在 V2 上解决率达 99%，而 Hard 子集将各系统区分在 52.9% 至 98.0% 之间。
正文节选
Scale AI Research \scaleailogopathassets/scale-logo-horizontal.png \partnerlogopathassets/reflection-logo.pdf \partnerlogoheight0.2in \contact{soham.dan, ying.liu, yue.zhang}@scale.com SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark Abstract SWE-Bench Pro is a widely used benchmark of long-horizon software engineering tasks for coding agents. As agents improve, they reach task issues earlier agents never touched and interact with the evaluation environment in ways the original harne