AI
介入的グラウンディング監査: 述語置換によるLLMのChain-of-Thoughtのブラックボックス前提条件テスト
Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution
arxiv2026年7月16日
日本語要約
本論文は、LLMのChain-of-Thoughtにおける前提条件依存性をテストするための新しいブラックボックス手法である介入的グラウンディング監査を提案する。述語を置換することで、内部状態に直接アクセスせずにLLMの推論をより詳細に分析およびデバッグできる。これはAIの解釈可能性と堅牢性を向上させる。
English Summary
The paper proposes Interventional Grounding Audits, a novel black-box method to test premise-dependency in LLM Chain-of-Thought. By substituting predicates, it allows for finer-grained analysis and debugging of LLM reasoning without direct access to internal states. This advances AI interpretability and robustness.