Beyond Prompt Engineering: Why Personalization in Explainable AI Requires Closed-Loop Control

The integration of Large Language Models into Explainable Artificial Intelligence has created a seductive trap. Faced with the complexity of technical explanation artifacts, practitioners increasingly rely on LLMs to translate feature attributions and counterfactuals into natural language narratives. The prevailing assumption appears to be that personalization simply requires better prompt engineering. Add demographic tags, adjust the system message for expertise levels, and the system will adapt to individual users. This approach, however, fundamentally misunderstands the nature of explanation as an interactive, epistemic process.

In their paper, PONTE: Personalized Orchestration for Natural Language Trustworthy Explanations, Vineis et al. challenge this assumption by reframing explanation generation not as a translation task but as a control problem. Their framework treats personalization as a closed-loop validation process where user preferences are modeled as a dynamic latent state rather than static annotations. This distinction carries profound implications for how we design trustworthy XAI systems in high-stakes domains such as healthcare and finance.

The Failure of Static Persona Assignment

Current approaches to personalizing XAI narratives typically rely on what might be called demographic prompt engineering. Systems are instructed to "explain like the user is a medical professional" or "use simple language for non-experts." This method treats personalization as a conditional generation task where user categories are discrete tokens prepended to prompts.

The PONTE framework identifies why this approach fails to achieve genuine adaptation. Static tags lack feedback mechanisms; they cannot verify whether the generated explanation actually aligns with the user's evolving understanding or specific informational needs. More critically, without grounding in structured XAI artifacts, LLMs generating these explanations may produce plausible-sounding but unfaithful narratives. The system is essentially hallucinating personas rather than adapting to real users.

PONTE addresses this by modeling personalization through a low-dimensional preference vector that captures stylistic and cognitive requirements. This is not merely a prompt modifier but a dynamic state updated through iterative user feedback. The framework explicitly separates the user's preference state from the generation process, creating a formal mechanism for tracking how explanation styles should evolve based on interaction history.

Verification as a Core Architectural Component

What distinguishes PONTE from conventional narrative generation systems is its integration of verification modules as first-class citizens rather than afterthoughts. The framework incorporates three specific validation layers:

• Numerical faithfulness, ensuring quantitative claims match the underlying XAI artifacts
• Informational completeness, verifying that all relevant explanation components are included
• Stylistic alignment, confirming the narrative matches the current preference state

This verification-refinement loop represents a fundamental architectural shift. Most LLM-based explanation systems generate text and hope for the best, perhaps with human evaluation during development. PONTE instead enforces deterministic constraints between the structured explanation artifacts (such as SHAP values or counterfactual rules) and the generated natural language. When the verification modules detect misalignment, the system refines the output rather than presenting potentially hallucinated content to the user.

The framework further incorporates retrieval-augmented generation to constrain explanatory claims to certified domain literature. This mitigates the risk of plausibility-masked hallucinations, where an LLM might invent domain-specific reasoning that sounds correct but lacks empirical grounding. By anchoring the narrative generation to verified external knowledge sources, PONTE ensures that even personalized explanations maintain epistemic integrity.

Empirical evaluations across healthcare and finance domains demonstrate the efficacy of this approach. Automatic metrics and human studies confirm that the verification-refinement loop substantially improves both completeness and stylistic alignment compared to validation-free generation. Crucially, human evaluators showed strong agreement between the system's intended preference vectors and the perceived style of explanations, suggesting that the low-dimensional preference model successfully captures meaningful stylistic dimensions.

Your Take

The PONTE framework illuminates a broader tension in modern XAI between accessibility and accuracy. As regulatory frameworks such as the GDPR and EU AI Act mandate transparency and human oversight, there is increasing pressure to make machine learning systems interpretable to non-technical stakeholders. However, simplification without verification risks creating what might be called explanatory theater: narratives that satisfy user expectations while misrepresenting the underlying model behavior.

My analysis suggests that the closed-loop approach proposed by Vineis et al. represents something more significant than an engineering improvement. It signals an epistemological shift from viewing explanation as a unidirectional information transmission (model to user) toward understanding it as a collaborative construction of meaning. The user is not a passive recipient of a pre-computed explanation but an active participant whose feedback continuously reshapes the explanation generation process.

However, this framework also raises important limitations and open questions. The human-in-the-loop design, while essential for preference learning, introduces scalability constraints. In high-throughput applications such as real-time credit scoring or emergency medical triage, the latency and cognitive load of iterative feedback loops may prove prohibitive. There is also the risk of preference collapse, where the system overfits to a user's stated preferences at the expense of exposing them to novel but relevant information that falls outside their