All Explanations are Wrong, But Many Are Useful: Exploring the Rashomon Explanation Set with Large Language Models

This paper introduces the Rashomon Explanation paradigm, which treats explanations as a set of faithful, prediction-guiding alternatives rather than a single canonical account. Its RashomonLLM workflow iteratively generates explanations, aligns predictions with them, and reflects on discrepancies. The authors claim improvements over state-of-the-art prediction and XAI baselines across customer-churn classification, clinical-survival regression, and industrial click-through prediction, with robustness to distribution shifts, temporal splits, and random seeds. The abstract also claims convergence and recovery of the full explanation set, but does not report the concrete effect sizes or experimental details.