The authors introduce the Rashomon Explanation paradigm, which builds a set of faithful, prediction-guiding explanations rather than relying on a single one. They propose RashomonLLM, an agentic workflow that generates natural language explanations by iteratively aligning them with predictions.
- The framework proves that explanation fidelity bounds model performance and that the set of explanations is generally non-empty.
- RashomonLLM converges and recovers the full set of explanations through its Explanation-Prediction-Reflection workflow.
- It significantly outperforms state-of-the-art prediction and XAI baselines on accuracy and explanation quality across customer-churn, clinical survival, and industrial click-through tasks.
The approach advances business performance while laying the groundwork for consumer trust by demonstrating that coupling explanation and prediction improves both.