📄 中文摘要
在使用大型语言模型（LLM）进行需要准确性的任务时，用户常常会发现其初始输出包含真实见解和自信的虚构内容。通过手动审查这些输出，用户可以识别并修正错误。然而，令人惊讶的是，当同一模型被要求审查其自身输出时，通常能够发现之前的错误，包括虚构的事实、逻辑漏洞和不一致之处。通过反复进行审查和修正，直到没有发现问题为止，最终结果会显得相当干净。这一过程虽然看似矛盾，但却有效地提高了输出的准确性。
迭代审查-修复循环消除大型语言模型的幻觉，并且有一个公式
📄 English Summary
Iterative review-fix loops remove LLM hallucinations, and there is a formula for it
When using large language models (LLMs) for tasks requiring accuracy, users often notice that the initial output is a mix of genuine insights and confident fabrications. By manually reviewing these outputs, users can identify and correct errors. Surprisingly, when the same model is asked to review its own output, it can often detect the very mistakes it made earlier, including hallucinated facts, logical gaps, and inconsistencies. This review-fix process can be repeated until no issues are found, resulting in remarkably clean outputs. Although this may seem paradoxical, it effectively enhances the accuracy of the results.