智能体与自主科学
突破级
暂无讲解视频
收录解读
Kintsugi: Learning Policies by Repairing Executable Knowledge Bases 关注的是一个可复用的 AI 系统或评测问题，而不是单点 demo。
Verifier-gated policy learning through editable executable knowledge bases.
It converts policy improvement into inspectable typed KB repair with deterministic execution, useful for embodied and long-horizon agents.
它没有更高，是因为这些新 arXiv 工作仍需要更多独立复现、真实系统部署和长期社区采用来确认影响。