RashomonLLM breaks accuracy-explainability trade-off with multi-explanation AI

A persistent belief in machine learning holds that making models explainable comes at the cost of accuracy. Pan Li's new paper challenges that dogma by introducing the Rashomon Explanation paradigm—a set of faithful, prediction-guiding explanations rather than a single explanation. The paper proves this set is non-empty and that explanation fidelity bounds model performance. To explore it, Li proposes RashomonLLM, an agentic workflow that uses an LLM to generate natural language explanations, then iteratively aligns them with predictions via reflection cycles. The method is proven to converge and recover the full explanation set.

RashomonLLM was tested across three domains: customer-churn classification, clinical survival regression, and industrial click-through prediction on large-scale live-streaming logs. It significantly outperformed state-of-the-art prediction and XAI baselines on both accuracy and explanation quality. Notably, the gains were robust to distribution shifts, temporal splits, and random seeds. This suggests that the traditional trade-off between explanation and accuracy is not fundamental—it arises from treating them as separate objectives. By coupling them, RashomonLLM offers a path to high-performing models that can also explain their reasoning, advancing both business outcomes and consumer trust.