Large language models (LLMs) are increasingly deployed for consequential decisions, making the pair of answer and explanation a crucial audit tool. Unfortunately, explanations can be unfaithful, failing to mirror the true reasoning. We distinguish two dimensions of unfaithfulness: incompleteness, where the explanation omits factors that affect the answer, and unsoundness, where it cites factors that did not influence the answer. Existing faithfulness‑enhancing methods fall into two groups: training‑time approaches that require access to model weights and heavy computation, and test‑time approaches that mainly address unsoundness. We propose a test‑time technique that directly tackles incompleteness: remove from the input all concepts not mentioned in the model’s explanation, then re‑query the model on this reduced input. This eliminates hidden influences while preserving the impact of the cited concepts. Experiments on two datasets, across several model families, and using two independent faithfulness metrics show that our method improves explanation faithfulness compared to standard prompting and prompting designed for faithfulness. The approach is model‑agnostic, applicable at inference without altering parameters, offering a flexible way to reduce hidden influences and enhance the reliability and safety of LLM‑assisted decision making.
Review