A Removal Based Approach to Improve LLM Faithfulness at Test-Time

One way to make an AI's explanation better match its decision is to restrict the information available for a second answer. A proposed method first asks a model for an answer and explanation, removes input concepts the explanation did not credit, and asks again using the reduced input. It targets omitted influences: information that affects a decision without appearing in its explanation. Across two datasets and multiple model families, the authors report improved explanation faithfulness under two metrics. The intervention changes what the model can use on the second pass, without retraining its parameters.