PONTE: Personalized Orchestration for Natural Language Trustworthy Explanations
| Authors | Vittoria Vineis et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2603.06485 |
| Download | |
| Categories | cs.CL, cs.AI |
Abstract
Explainable Artificial Intelligence (XAI) seeks to enhance the transparency and accountability of machine learning systems, yet most methods follow a one-size-fits-all paradigm that neglects user differences in expertise, goals, and cognitive needs. Although Large Language Models can translate technical explanations into natural language, they introduce challenges related to faithfulness and hallucinations. To address these challenges, we present PONTE (Personalized Orchestration for Natural language Trustworthy Explanations), a human-in-the-loop framework for adaptive and reliable XAI narratives. PONTE models personalization as a closed-loop validation and adaptation process rather than prompt engineering. It combines: (i) a low-dimensional preference model capturing stylistic requirements; (ii) a preference-conditioned generator grounded in structured XAI artifacts; and (iii) verification modules enforcing numerical faithfulness, informational completeness, and stylistic alignment, optionally supported by retrieval-grounded argumentation. User feedback iteratively updates the preference state, enabling quick personalization. Automatic and human evaluations across healthcare and finance domains show that the verification-refinement loop substantially improves completeness and stylistic alignment over validation-free generation. Human studies further confirm strong agreement between intended preference vectors and perceived style, robustness to generation stochasticity, and consistently positive quality assessments.
Engineering Breakdown
Plain English
PONTE is a framework that generates personalized explanations for ML model outputs using Large Language Models, but instead of just prompt engineering, it treats personalization as a closed-loop validation process. The paper addresses a real problem: current XAI methods ignore that users have different expertise levels, goals, and cognitive needs—a domain expert needs different explanations than a non-technical stakeholder. PONTE combines a low-dimensional preference model (capturing stylistic requirements) with human-in-the-loop feedback to produce trustworthy, adapted narratives that avoid hallucinations and maintain faithfulness to the underlying model's decisions. The framework moves beyond one-size-fits-all explanation pipelines by continuously validating and adapting outputs based on user interaction.
Core Technical Contribution
The core novelty is modeling personalization as an iterative validation-adaptation loop rather than a static prompt engineering problem. Instead of fine-tuning prompts upfront to match a user profile, PONTE builds a low-dimensional preference model that captures user-specific stylistic requirements and continuously validates LLM outputs against ground-truth model behavior before presenting them. This is fundamentally different from prior work that either treats all users identically or relies on prompt variations without verification. The closed-loop human-in-the-loop design ensures explanations remain faithful to the underlying model while adapting presentation style—a key technical distinction that separates PONTE from prompt-based personalization or static explanation templates.
How It Works
PONTE operates in three main phases: (1) profile construction, where the system learns a low-dimensional representation of each user's explanation preferences (e.g., technical depth, verbosity, example types) through initial interaction or explicit input, (2) explanation generation, where an LLM produces candidate narratives conditioned on this preference model alongside the model output being explained, and (3) validation-adaptation, where the system checks generated explanations against ground-truth model behavior (via feature importance, activation patterns, or decision rules) and flags hallucinations or inconsistencies before returning to the user. If validation fails, the framework can either refine the preference model based on the mismatch, regenerate the explanation with adjusted constraints, or request human feedback to recalibrate. The low-dimensional preference model allows efficient personalization without per-user fine-tuning, while the validation loop prevents the LLM from producing plausible-sounding but incorrect explanations—a critical failure mode in LLM-based XAI.
Production Impact
In production, PONTE solves three concrete problems: (1) serving explanations to heterogeneous users (regulators, domain experts, end-users) without maintaining separate pipelines, (2) reducing the cost and risk of LLM hallucinations in safety-critical domains (finance, healthcare, legal) where incorrect explanations can mislead decisions, and (3) scaling personalized XAI without manual engineering per user segment. The trade-offs are meaningful: each user interaction adds a validation round (latency impact depends on validation method—simple feature checks are fast; complex behavioral validation adds seconds), the preference model requires initial calibration data per user or user segment, and the framework assumes access to ground-truth model behavior (weights, activations) which may not exist for black-box third-party models. For a typical deployment, expect 20-40% latency overhead for the validation loop and need to store/maintain a preference embedding per user or user cohort. Integration complexity is moderate—the framework wraps existing LLM APIs and explanation methods but requires defining what 'faithful' means for your specific model.
Limitations and When Not to Use This
PONTE's approach is constrained by several critical assumptions that may not hold in production: (1) it assumes you have access to ground-truth model behavior for validation, which fails for proprietary or ensemble models where internals are opaque, (2) the low-dimensional preference model assumes user preferences are stable and can be captured in relatively few dimensions—real users have context-dependent, evolving explanation needs that this model may oversimplify, (3) the framework doesn't address what happens when the base model itself is biased or unfaithful, as it can only validate explanations against model behavior, not ground truth, and (4) the paper doesn't clearly specify how to handle adversarial users or preference-drift over time. Additionally, the approach is computationally more expensive than simple prompt-based personalization because it adds a validation loop, and it doesn't solve the fundamental problem that LLMs sometimes generate coherent but false explanations—it only catches inconsistencies with model internals, not real-world inaccuracies. Follow-up work needs to address cross-domain preference transfer, validation strategies for black-box models, and how to detect when the base model's behavior itself is untrustworthy.
Research Context
PONTE builds on two converging research threads: (1) adaptive XAI systems that recognize users have different needs (prior work like Trepan, LIME variants adapted for user types), and (2) the emerging challenge of controlling LLM reliability in explanation tasks (related to work on hallucination detection, constitutional AI, and LLM grounding). The paper sits at the intersection of human-centered AI (HCI + XAI) and LLM safety/alignment, addressing a gap where prior explanation methods were either generic or required manual prompt engineering per user. It advances beyond one-off user studies by proposing a generalizable framework that can work across different base models and domains, positioning it as a practical contribution to the XAI toolbox rather than a purely theoretical advance. The research opens directions around user preference modeling for AI explanations, automated validation strategies for LLM outputs, and the broader question of how to serve personalized explanations at scale without sacrificing truthfulness.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::