Hacker News 热门(buzzing.cc 中文翻译)📅 Jul 2AI 产品

Senior SWE-Bench:评估AI智能体作为高级工程师的基准测试

Senior SWE-Bench是一个开源基准测试,用于评估AI智能体完成高级软件工程师级别任务的能力。任务分功能开发与Bug修复两类:功能任务指令类似自然语言消息,采用验证智能体基于专家配方自动生成行为测试;Bug任务要求根据日志、profiling等运行时信息深入调查。排行榜显示,Claude Opus 4.8搭配Mini-SWE-Agent(max effort)通过率24.0%,Claude Sonnet 5为19.4%,GPT-5.5为16.0%,最强前沿模型在超75%任务中未能达到高级工程师级别的正确性与品味。每个功能任务平均涉及11个文件,最强智能体也需数百步完成;中位指令长度仅为SWE-Bench Pro的31%。任务来源于从库到多服务应用的仓库PR,由拥有数百次提交的工程师编写。

💡 Recommended AI Tools

A comprehensive Chinese base model evaluation suite
The large video model released by Shengshu Technology and Tsinghua University supports highly consistent simulation of characters and physical laws.
The large video generation model launched by Kuaishou supports the generation of 1080p frame rate videos up to 2 minutes long.
ByteDance's AI painting and short video generation platform supports one-click text generation and shadow generation.