Ask HN：如何应对过度“做好事”的代理人
2 分•作者: zof3•大约 1 个月前
您好 HN-
我们团队最近对使用的每一个大型语言模型都感到厌烦，因为即使是处理简单的事情，它们也会默认发出公司合规官或循规蹈矩者的声音（例如，“嗯，但是 XYZ”等等）。我们得出的结论是，训练后进行的调整过于严格，导致模型默认表现得像个傻瓜。我们创建了一些内部的 agents.md 文件和预提示指令来尝试打破这种模式——有时是有效的——但你们大家是如何应对或绕过这个问题的呢？
查看原文
Hi HN-<p>Our team has been recently getting fed up with every single LLM that we use defaulting to sounding like a corporate compliance officer or goody-two-shoes (‘erm, but xyz’ etc) on even simple things. We’ve come to the conclusion that the post training being done is so strict that models default to acting like Gomer Pyle. We’ve made some internal agents.md files and pre-prompt instructions to try to kick them out of the pattern — and it works sometimes — but how are all of you dealing with this or working around it?