The Problem: Illusory Chain-of-Thought
Large language models often produce reasoning chains that appear logically sound but fail to genuinely depend on the provided premises. Current evaluation methods, such as self-consistency, often fail to detect when a model arrives at the correct conclusion through flawed or hallucinated reasoning paths.
The Solution: Interventional Grounding Audits
Interventional grounding audits provide a black-box, step-level test to verify premise dependency. The process works by:
- Targeting a premise: Selecting a specific premise within a reasoning chain.
- Predicate Substitution: Intervening on that premise by substituting its target predicate with a fresh, novel symbol.
- Re-evaluation: Re-running the model with the modified input.
- Dependency Check: Observing whether the conclusion of each reasoning step changes. If the conclusion remains unchanged despite the substitution, the model is not actually grounding its reasoning in that specific premise.
Performance and Insights
When tested on the ProntoQA benchmark using GPT-4o, this method achieved an F1 score of 0.806 in detecting proof-tree dependencies, significantly outperforming the self-consistency baseline (F1 = 0.343).
Crucially, the audit revealed that 66% of correctly solved problems contained at least one step that was insensitive to its required premise. These failures were primarily linked to 'entity-introduction' premises, which the researchers identified as a blind spot for consistent-substitution evaluators. This highlights a critical 'right answer, wrong reasoning' signal that passive evaluation methods consistently miss.