A framework (FITCF, building on ZEROCF) that generates textual counterfactual examples by first extracting the most important words of an input via feature-attribution methods (LIME, SHAP, Integrated Gradients) on a fine-tuned classifier, then prompting an LLM zero-shot with those words, verifying each candidate with a label-flip check, and reinserting the verified counterfactuals as few-shot demonstrations. The key empirical finding is that the faithfulness of the feature-attribution scores correlates strongly with the quality of the generated counterfactuals, and that the number of demonstrations is the dominant performance factor.
Introduces ZEROCF, a zero-shot method that generates counterfactual examples by prompting LLMs with the most important words identified via feature-attribution methods (LIME, SHAP, Integrated Gradien…