New audit method finds majority of LLM chain-of-thought steps insensitive to premise dependencies

Researchers from arXiv cs.AI introduce interventional grounding audits, a black-box method to test whether individual steps in an LLM's chain-of-thought (CoT) reasoning genuinely depend on stated premises. The approach intervenes by substituting a premise's target predicate with a fresh symbol, re-runs the model, and checks whether the normalized conclusion of each reasoning step changes. This step-level intervention contrasts with passive evaluation methods that do not actively probe premise dependency.

The method was evaluated on ProntoQA, a synthetic multi-hop deductive reasoning benchmark with gold proof trees, where step-level premise dependencies are known. When applied to 50 ProntoQA problems using GPT-4o, the interventional grounding audits achieved an F1 score of 0.806 for detecting proof-tree dependencies. Precision and recall for predicate-determining dependencies were 0.885 and 1.00, respectively, significantly outperforming a self-consistency baseline, which recorded an F1 of 0.343 with non-overlapping 95% bootstrap confidence intervals.

The study further identifies that 66% of correctly solved problems contained at least one aligned reasoning step that was insensitive to a direct proof-tree dependency under consistent substitution. These errors were exclusively associated with entity-introduction premises, a documented blind spot of consistent-substitution evaluators. The authors describe this as a 'right answer, wrong reasoning' signal that remains invisible to passive evaluation methods.

All audit certificates, raw outputs, and reproduction scripts are publicly available in a GitHub repository. The paper also discusses scope limitations, noting that the method's applicability may be constrained to formal, parsable benchmarks and may not generalize to open-ended or unstructured reasoning tasks.