TL;DR
- 通过让 LLM 生成 feedback 不断 refine 自身的回答，可以提升回答效果，但也会带来不可忽视的推理开销
![[NeurIPS 2024] Self-Refine: Iterative Refinement with Self-Feedback,在这里插入图片描述,第1张](https://i-blog.csdnimg.cn/direct/250fb0573515427aa5d9faec6d4258e5.png#pic_center "[NeurIPS 2024] Self-Refine: Iterative Refinement with Self-Feedback,第1张")
![[NeurIPS 2024] Self-Refine: Iterative Refinement with Self-Feedback,在这里插入图片描述,第2张](https://i-blog.csdnimg.cn/direct/710d7b72a02e44e4a7c5a568ceda1ce0.png#pic_center "[NeurIPS 2024] Self-Refine: Iterative Refinement with Self-Feedback,第2张")
![[NeurIPS 2024] Self-Refine: Iterative Refinement with Self-Feedback,在这里插入图片描述,第3张](https://i-blog.csdnimg.cn/direct/f82aaf029cb840f49bd520e185499194.png#pic_center "[NeurIPS 2024] Self-Refine: Iterative Refinement with Self-Feedback,第3张")
References
- Madaan, Aman, et al. “Self-refine: Iterative refinement with self-feedback.” Advances in Neural Information Processing Systems 36 (2024).
- code: https://selfrefine.info/
本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明原文出处。如若内容造成侵权/违法违规/事实不符，请联系SD编程学习网：675289112@qq.com进行投诉反馈，一经查实，立即删除！
