arXiv Artificial Intelligence· Michael Kirchhof, Eleonora Gualdoni, Andrew Szot, Khashayar Gatmiry, Aryo Lotfi, Abbas Kazerouni, Omar Attia, Sanjoy Chowdhury, Alexander Toshev·· 3 小时前AI 评分38
RLTLDR：通过内化自生成反馈实现自我提升
RLTLDR: Self-improvement by Internalizing Self-generated Feedback
AI 导读
RLTLDR 让策略在每次失败后读取验证器输出并自行写出一条 TL;DR 式洞察，后续 rollout 以历史洞察为条件，并对上下文中的洞察做反向传播以内化任务到洞察的映射。
来源：arXiv Artificial Intelligence · arxiv.org