HackerNews中文版

一个表单同时完成了五项任务：判断状态、选择要打开的表单、提取数值、记录每个数值的来源，以及验证必填字段。自然语言保留了第三项（提取数值），其余的交给了模型。没有人记录哪些信息丢失了。大型语言模型（LLMs）是通过填空来训练的。现在我们已经从对话转向了行动，我们却告诉它们不要填空。但是，没有人给它们一份不允许推断的内容列表，也没有教它们如何在不推断的情况下填空。这里是对该列表和该方法的尝试。 https://github.com/Jang-woo-AnnaSoft/execution-state-preflight/

查看原文

A form did five things at once: judged the condition, picked which form to open, took the values, carried where each value came from, and validated the required fields.Natural language kept the third one and gave the rest to the model. Nobody wrote down which ones went missing.LLMs were trained by filling in blanks. Now that we've moved from conversation to action, we tell them not to fill in blanks.But nobody has handed them a list of what they aren't allowed to infer, or shown them how to fill a blank without inferring.Here's an attempt at that list and that method.https://github.com/Jang-woo-AnnaSoft/execution-state-preflight/

表格完成了五项功能，自然语言保留了一项。