PONTE: Personalized Orchestration for Natural Language Trustworthy Explanations

One-size-fits-all explanations don't work: PONTE addresses a critical gap in current AI explainability tools by recognizing that different users—whether experts or non-technical stakeholders—have different needs, expertise levels, and cognitive requirements for understanding AI decisions.

AI explanations can be unreliable: The paper identifies a key problem with using Large Language Models to explain AI systems: they can introduce "hallucinations" (false information) and lack faithfulness to the actual technical explanations, undermining trust in the explanations themselves.

Verification is essential for trustworthy explanations: PONTE's core innovation is a verification-refinement loop that checks explanations for numerical accuracy, completeness of information, and stylistic appropriateness—rather than relying solely on prompt engineering—which human evaluations showed "substantially improves" explanation quality.

User feedback enables better personalization: Rather than one-time customization, PONTE uses an iterative, human-in-the-loop approach where user feedback continuously updates how explanations are tailored, allowing systems to quickly adapt to individual preferences.

Personalized explanations improve user understanding: Testing in healthcare and finance domains demonstrated that explanations personalized through this verification process achieved "strong agreement between intended preference vectors and perceived style," suggesting users actually receive explanations matched to their needs.