Large language models (LLMs) are used for decisions that require trustworthy explanations. These explanations can be unfaithful, either omitting relevant factors (incompleteness) or citing irrelevant ones (unsoundness).
Existing methods for improving faithfulness often involve training-time adjustments or focus on unsoundness at test time. The introduced approach targets incompleteness by removing concepts not credited in the explanation from the input and re-querying the model.
This process reduces hidden influences and maintains the influence of credited concepts, leading to more faithful explanations. It is model-agnostic and can be applied during inference without altering model parameters.
The approach was tested across multiple datasets, model families, and faithfulness metrics, showing improvements over standard prompting methods.
Source: https://arxiv.org/abs/2609.04343