V2LLM: A Digital-Twin-Aware Retrieval-Augmented Large Language Model for Resilient Telemetry Recovery and Fault Diagnosis in V2X

Reliable telemetry is essential for fault detection and adaptive control in vehicle-to-everything (V2X) networks, yet wireless impairments often cause bursty, non-i.i.d. data loss. This paper proposes V2LLM, a digital-twin-aware, retrieval-augmented large language model (LLM) framework for resilient telemetry recovery and continual fault diagnosis. Leveraging the reasoning capabilities of large AI models (LAMs), V2LLM performs token-efficient reconstruction by combining local history, memory-guided context, and digital twin predictions. A sliding-window autoregressive scheme enables structured recovery under long gaps, while a federated continual learning module adapts fault classifiers at roadside units without raw data sharing. V2LLM tightly integrates semantic prompting, predictive modeling, and distributed adaptation to address uncertainty, heterogeneity, and spatiotemporal drift in V2X environments. Extensive evaluations show that the proposed framework consistently improves reconstruction quality and classification robustness, demonstrating the potential of LAM-powered architectures for intelligent, resilient wireless systems.