---
id: 20260717-T0-15
title: "黑盒干预审计：检测LLM思维链是否真的依赖前提"
title_en: "Black-Box Audit Tests Whether LLM Chain-of-Thought Relies on Stated Premises"
url: https://ai.daily.yangsir.net/daily/20260717-T0-15
issue_date: 2026-07-17
publish_date: 2026-07-16T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2607.13069
---

# 黑盒干预审计：检测LLM思维链是否真的依赖前提

一项新研究提出“干预式基础审计”（Interventional Grounding Audits），这是一种黑盒、步骤级的测试方法，通过替换推理中的谓词来检测大语言模型（LLM）的思维链（CoT）是否真正依赖其陈述的前提。实验发现，许多看似逻辑严密的CoT推理实际上存在“前提漂移”——模型并未真正使用它声称要依赖的假设。该工具可帮助开发者评估LLM推理的可信度与一致性。

## English Version

**Black-Box Audit Tests Whether LLM Chain-of-Thought Relies on Stated Premises**

A new study introduces Interventional Grounding Audits, a black-box, step-level test that checks whether an LLM's chain-of-thought (CoT) genuinely depends on its stated premises by substituting predicates during reasoning. Experiments reveal that many seemingly logical CoT chains suffer from 'premise drift'—the model does not actually use the assumptions it claims to rely on. The tool helps developers evaluate the trustworthiness and consistency of LLM reasoning.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2607.13069)

**详情页**：https://ai.daily.yangsir.net/daily/20260717-T0-15

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*