大型语言模型学习的是程序员创造的内容，而不是程序员的工作方式。
2 分•作者: noemit•5 个月前
我做了一个实验，想看看 CLI 到底是不是调用工具最直观的格式。（正如一位前 Manus AI 后端工程师所声称的那样）我给我的模型提供了随机场景和一个名为“run”的单一工具——我告诉它这个工具的工作方式类似于 CLI。我让它猜测命令。
它猜测的命令很棒，但它总是以冒号开头来格式化，例如：
:help
:browser
:search
:curl
它接受的训练是终端的样子，而不是你实际输入的内容（你不会输入“:”）。
从那以后，我更新了我的代理工具中的代码，不再与这种直觉对抗。
大型语言模型学习的是文档/制品中命令的样子，而不是人类在键盘上实际输入的内容。
这似乎很明显。这就是为什么你必须测试你的大型语言模型，看看它自然是如何工作的，这样你就不用用你的系统提示来与它对抗了。
顺便说一句，这是 Kimi K2.5。
查看原文
I ran an experiment to see if CLI actually was the most intuitive format for tool calling. (As claimed by a ex-Manus AI Backend Engineer) I gave my model random scenarios and a single tool "run" - i told it that it worked like a CLI. I told it to guess commands.<p>it guessed great commands, but it formatted it always with a colon up front, like
:help
:browser
:search
:curl<p>It was trained on how terminals look, not what you actually type (you don't type the ":")<p>I have since updated my code in my agent tool to stop fighting against this intuition.<p>LLMs they learn what commands look like in documentation/artifacts, not what the human actually typed on the keyboard.<p>Seems so obvious. This is why you have to test your LLM and see how it naturally works, so you don't have to fight it with your system prompt.<p>This is Kimi K2.5 Btw.