New Method Audits LLM Reasoning for Premise Dependency

Large language models (LLMs) often generate reasoning chains that appear logically sound, yet it's not always clear if their conclusions truly stem from the premises they cite. To address this, a new technique called "interventional grounding audits" has been developed. This black-box, step-level testing method assesses premise dependency by systematically altering a single premise through predicate substitution and then observing how the LLM's reasoning steps and normalized conclusions change. The method was evaluated on ProntoQA, a synthetic benchmark for multi-hop deductive reasoning where true premise dependencies are known. When applied to GPT-4o, the interventional grounding audits achieved a high F1 score of 0.806 in detecting proof-tree dependencies, significantly outperforming a self-consistency baseline. This indicates a more robust way to verify the integrity of an LLM's reasoning process. Furthermore, the research revealed that a notable portion of correctly solved problems still contained reasoning steps insensitive to direct proof-tree dependencies, particularly involving entity-introduction premises. This highlights a "right answer, wrong reasoning" phenomenon that passive evaluation methods often miss. The findings suggest that while LLMs can arrive at correct answers, their internal reasoning might not always align with human-expected logical dependencies, emphasizing the need for more rigorous auditing tools.