Snowflake 研究：近三分之一软件工程训练任务会接受"通过测试但未修复 bug"的补丁，推出 Arctic-SWE
What happens when a coding agent gets rewarded for a patch that passes tests but doesn’t fix the bug? In Snowflake AI Research team's audit of three public software-engineering training datasets, nearly a third of standard-validated tasks accepted a plausible wrong patch. Arctic-SWE combines a stress-tested reinforcement learning recipe with curated tasks and stronger tests. Explore the failure modes and the fixes: https://t.co/hNZS33Q0CF
Snowflake AI Research 团队审计三个公开软件工程训练数据集后发现，近三分之一经标准验证的任务会接受看似合理但实际未修复 bug 的补丁。为此团队推出 Arctic-SWE，将经过压力测试的强化学习方案与精选任务、更严格的测试相结合。
粉丝 6.6万
The AI Data Cloud where data does more, and proud to be the Official Data Collaboration Provider for LA28 and Team USA.
- 浏览
- 2049
- 点赞
- 17
- 评论
- 5
- 分享
- 2
- 收藏
- 7
约为这个来源平时的 2.3 倍
数字来自平台，为最近一次抓取时的数据。
来源：X · Snowflake · x.com