EECA: a design-oriented explainability framework for end-to-end cockpit agents

As automotive cockpits evolve into proactive interaction environments supported by End-to-End (E2E) Large Language Models (LLMs), their opaque black-box reasoning processes may create a sensing–design disconnect, contributing to automation surprise and reduced human–AI trust. To address this transparency challenge, this study repositions Explainable AI (XAI) from a post-hoc diagnostic tool toward a design-oriented methodology. We propose the Explainable End-to-End Cockpit Agent (EECA) framework, which integrates multi-sensor fusion, LLM-based reasoning, and structured graph logic. An Explainable Interaction Design Knowledge Graph further maps multimodal driver states to experience goals, design attributes, explanation cues, and evaluation indicators, thereby supporting the generation of transparent and explainable multimodal interaction prompts. A hardware–software prototype was evaluated across 200 complex cockpit scenarios. Compared with selected baseline models, EECA achieved an F1-score of 0.892 and an information gain of 8.3 under edge-side inference conditions. Furthermore, a simulator-based study with 30 participants provided preliminary mixed-method evidence that EECA was associated with shorter driver takeover reaction time and higher trust-related user-experience ratings under the tested cockpit scenarios.