A Unified Framework with Novel Metrics for Evaluating the Effectiveness of XAI Techniques in LLMs

Different explanation methods have different strengths: The study found that no single XAI technique excels at everything—LIME performed consistently well across most measures, AMV (Attention Mechanism Visualization) was most robust to small input changes, and LRP was best at highlighting contrastive differences between predictions. This suggests organizations should match their choice of explanation method to their specific needs.

Explanation quality can be measured systematically: The research introduced four concrete metrics for evaluating XAI techniques—Human-reasoning Agreement (whether explanations match how humans think), Robustness (stability when inputs change slightly), Consistency (reliability across similar cases), and Contrastivity (ability to show why one prediction differs from another). These metrics provide a standardized way to assess whether explanations are actually trustworthy.

Simpler explanation methods may be more reliable: LIME, which is a relatively straightforward technique that approximates model behavior locally, consistently outperformed more complex methods across multiple AI models and tasks. This finding suggests that transparency and trustworthiness don't necessarily require the most sophisticated technical approaches.

Explanation effectiveness varies by model complexity: The study tested five different LLMs and found that explanation quality differed depending on which model was being explained. This indicates that organizations cannot assume one explanation method will work equally well across all their AI systems—evaluation and testing are necessary for each deployment.