Zero-Shot Explainable AI: Bridging the Gap Between Prior-Fitted Networks and Human-Centric Interpretability

Abstract

Machine learning on tabular data has historically relied on computationally expensive pipelines, involving extensive hyperparameter tuning and feature engineering. Furthermore, the highest-performing models, such as gradient boosting machines and deep neural networks, often operate as opaque "black boxes," severely limiting their deployment in high-stakes domains that require transparency. This study presents a novel end-to-end framework integrating TabPFN a transformer-based Prior-Fitted Network capable of zero-shot inference with SHapley Additive exPlanations (SHAP) and a Large Language Model (LLM) narrative engine. The purpose of this research is to evaluate whether a zero-shot architecture can achieve competitive predictive performance while simultaneously offering automated, human-readable interpretability. The methodology involves constructing an interactive microservices-based dashboard that processes raw tabular data, executes GPU-accelerated inference, and dynamically translates statistical SHAP values into contextual text summaries. Findings indicate that the proposed system reduces end-to-end inference and explanation latency while completely eliminating the traditional training phase. The integration of LLM-generated narratives significantly lowers the barrier to entry for non-technical stakeholders, establishing a new paradigm for scalable, trustworthy machine learning on tabular data.