A Removal-Based Method Improves LLM Explanation Faithfulness at Test Time

Summary

Large language models may produce explanations that do not faithfully reflect the factors behind their answers. This paper distinguishes incompleteness, where an explanation omits influential factors, from unsoundness, where it cites factors that did not influence the answer. The authors introduce a model-agnostic test-time method aimed at incompleteness: concepts absent from the explanation are removed from the input, and the model is queried again on the reduced input. Across two datasets, multiple model families, and two independent faithfulness metrics, the method improves explanation faithfulness over both standard prompting and prompts that explicitly encourage faithfulness. It requires no model-parameter changes and is intended to reduce hidden influences in LLM-assisted decision making.