Ask HN：如何开发更具确定性的 LLM 管道？
1 分•作者: sky2224•大约 2 个月前
在使用大型语言模型（LLM）时，我通常会构建一个提示，然后得到一个结构不太好、需要人工审核才能验证结果质量的输出。而LLM的非确定性是导致这种情况的主要原因。
我看到一些使用LLM、视觉语言模型（VLM）等技术的解决方案，它们能够完成之前已经解决的问题（例如OCR、解析，甚至代码生成）。它们能很好、很快地解决这些问题……但通常只能达到90%的准确率。我该如何搭建一个框架，让我能够对我的解决方案有100%的信心（即100%确信我 *知道* 它会做什么）？
在传统的机器学习中，我可以通过置信度分数来判断是否接受一个输出。而且，通常情况下，它们是确定性的（即给定相同的输入，我总是会得到相同的输出）。而对于LLM等模型，我感觉这方面是缺失的。
查看原文
It feels like with LLMs I develop a prompt and then get some kind of output that's not very well structured and requires some kind of human oversight to verify that the result I've gotten back is of quality, and the non-deterministic nature of LLMs is the main reason for this.<p>I see solutions using LLMs, VLMs, etc that will achieve things that have been solved before (things like OCR, parsing, or even just code generation). They do these problems really well and really fast... 90% of the time. How do I get a scaffolding setup so I can be 100% confident in my solution (meaning 100% confident that I <i>know</i> what it's going to do)?<p>With standard ML, I have things like confidence scores to base whether or not I accept an output as valuable. And generally speaking, it's pure (i.e., given the same input I will always get the same output). With LLMs and the like, it feels like this piece is missing.