正在补充深度解读，当前内容可以先阅读
论文解决了什么问题
When a coding agent obeys a rule, it may simply have been going to do that anyway. Existing instruction-following benchmarks cannot tell the difference: they concentrate rules in the user turn, while coding-agent benchmarks emphasize final task success. We introduce Harness-IF, which scores operational rules one at a t...
适合谁阅读
上下文工程Agent 系统技能学习强化学习
可核验的原论文来源和作者
- 作者
- 作者信息暂未从原始元数据中确认
- 来源
- arXiv
- 论文 ID
- 2608.11727