An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
LinkedIn’s first prompt update from a recursive improvement loop depended on teaching the evaluators to distinguish real failures from ambiguous rubrics and bad diagnoses.
📌 One-Sentence Summary LinkedIn 用递归改进流程生成了首个上线的提示词更新。为了让改动靠谱,团队先修正了评估标准和归因方式。 📝 Summary Ziwei Guo 和 Saloni Gandhi 介绍了改进用户理解智能体的七个步骤,从收集执行轨迹,到提出修改、验证候选项,再用留出集测试。第一层评估器给输出打分,第二层诊断失败原因,再据此提出改动。团队复核约 180
来源:BestBlogs·AI 高分精选 · www.bestblogs.dev
LinkedIn’s first prompt update from a recursive improvement loop depended on teaching the evaluators to distinguish real failures from ambiguous rubrics and bad diagnoses.