---
schema_version: '1.0'
id: news-20260717-e4acd4
url: https://gotosocial.chinng-lab-srv.dev/interventional-grounding-audits-black-box-premise-dependency-tests-for-llm-chain-of-thought-via-predicate-substitution/
url_hash: e4acd49525d7e27a731fa2627aedc5fb4ac05a5efb1f1f4978bd7a6fa341816d
canonical_url: https://gotosocial.chinng-lab-srv.dev/interventional-grounding-audits-black-box-premise-dependency-tests-for-llm-chain-of-thought-via-predicate-substitution
source: ghost-chinng-lab
category: news/tech
category_raw: AI・テクノロジー
region: JP
tags:
- AI・テクノロジー
- ボックス評価
- Grounding
- 外部監査
- リスク管理
- Tests
lang: ja
published_at: '2026-07-17T00:25:08Z'
fetched_at: '2026-07-17T03:38:56Z'
updated_at: '2026-07-17T03:39:06Z'
status: published
content_hash: null
license_note: full
summary: LLMのChain-of-Thought推論において、前提がどの程度結論に依存しているかを測定する「介入的grounding audits」手法の研究。述語を新規シンボルに置換して推論を再実行し、前後での結果の変化を比較することで、推論過程の信頼性を評価。従来の自己一貫性ベースの評価では見落とされがちな前提依存性を直接測定でき、66%以上のケースで推論の乖離を検出できた。
summary_source: llm
summary_en: In LLM’s Chain-of-Thought reasoning, a study of the “inter。al grounding
  audits” method that defines how much conclusion the premise depends on. Review the
  reliability of the inference process by subst ting the predicate to a new symbol,
  and comparing the results before and after. In conventional self-consistency-based
  evaluations, we could directly measure the prerequisite dependencies that are often
  overlooked, and we could detect the divergence of reasoning in more than 66% cases。
entities:
- name: interventional grounding audits
  type: concept
- name: リスク管理
  type: concept
- name: '#protests'
  type: concept
key_facts: []
related: []
related_auto:
- name: Explainable AI
  type: concept
  weight: 1.0
- name: ブラックボックス
  type: concept
  weight: 1.0
- name: 投資戦略
  type: method
  weight: 1.0
- name: 投資家
  type: person
  weight: 1.0
- name: '#list'
  type: concept
  weight: 1.0
title: 'Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM
  Chain-of-Thought via Predicate Substitution'
---

# Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution

## TL;DR
LLMのChain-of-Thought推論において、前提がどの程度結論に依存しているかを測定する「介入的grounding audits」手法の研究。述語を新規シンボルに置換して推論を再実行し、前後での結果の変化を比較することで、推論過程の信頼性を評価。従来の自己一貫性ベースの評価では見落とされがちな前提依存性を直接測定でき、66%以上のケースで推論の乖離を検出できた。

## Key Points
- AI・テクノロジー / ボックス評価 / Grounding / 外部監査 / リスク管理 / Tests

## Details
(本文なし。リンク先参照)

## Source
元記事: [Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution](https://gotosocial.chinng-lab-srv.dev/interventional-grounding-audits-black-box-premise-dependency-tests-for-llm-chain-of-thought-via-predicate-substitution/) — published 2026-07-17T00:25:08Z