LLM-Guided Framework for Human-Centered XAI Selection in TinyML Edge Devices

A new framework for selecting explainable AI (XAI) methods in TinyML edge deployments has been proposed, addressing the challenge of balancing stakeholder preferences, explanation quality, and deployment costs. The framework, detailed in a preprint on arXiv (2608.07091), integrates a large language model (LLM)-guided design interface that translates qualitative stakeholder preferences into candidate XAI methods. This is followed by deterministic feasibility filtering and Pareto-based optimization to identify optimal solutions. The approach is particularly relevant for clinical applications where local and timely inference is critical, and where XAI serves as a human-AI interface to support healthcare professionals and patients in understanding model predictions. The work highlights the importance of human-centered design in AI deployment, ensuring that technical constraints and user needs are jointly considered. The framework is presented as a multi-objective design problem, emphasizing the trade-offs inherent in deploying AI on resource-constrained devices. The preprint was announced as a cross-type submission, indicating its interdisciplinary nature.