04:00
2026-07-16
arxiv.org
large-language-models
Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution
A new black-box auditing method called interventional grounding audits achieves F1=0.806 in detecting premise dependency in GPT-4o chain-of-thought reasoning on the ProntoQA benchmark, significantly o…