A Two-Stage LLM Framework for Accessible and Verified XAI Explanations

Verification is essential for trustworthy AI explanations: The paper demonstrates that without a verification step, LLM-generated explanations of AI decisions can contain inaccuracies and "hallucinations" (false information). The proposed two-stage framework filters out unreliable explanations before they reach users, addressing a critical gap in current approaches that lack safeguards.

Current explanation methods lack accountability: Existing approaches to translating technical AI outputs into plain language explanations "often lack guarantees of accuracy, faithfulness, and completeness," and evaluation remains "largely subjective or confined to post-hoc scoring." This means flawed explanations can currently reach end-users without detection.

Iterative refinement improves explanation quality: The framework uses feedback from a "Verifier LLM" to iteratively improve explanations, with analysis showing this process creates "more stable and coherent reasoning" over time. This suggests that AI-generated explanations benefit from a built-in quality control loop rather than a single-pass approach.

Accessibility and accuracy can be balanced: The research shows that natural-language explanations generated by LLMs are more linguistically accessible than raw technical outputs while maintaining reliability through verification—suggesting that making AI decisions understandable to non-experts doesn't require sacrificing accuracy.

Framework tested across multiple AI techniques: Experiments covered five different XAI methods and three families of open-weight LLMs, suggesting the approach is generalizable rather than limited to specific AI systems, which strengthens its potential real-world applicability.