Dynamic task offloading in vehicular networks using large language models for adaptive low latency decision making

Task offloading in vehicular environments is essential for efficient computation and resource utilization among connected vehicles. However, traditional approaches e.g., Deep Reinforcement Learning (DRL) and heuristic methods often struggle with dynamic adaptation, communication overhead, and scalability in dense, fast-changing scenarios. This paper proposes an edge-intelligent framework that leverages a Large Language Model (LLM) deployed at Roadside Unit (RSU) edge nodes to optimize dynamic, multi objective offloading decisions. The LLM is fine tuned on a structured dataset encoding real time vehicular states (mobility, CPU, bandwidth, battery), task characteristics, and historical offloading outcomes, enabling reasoning over multi-dimensional inputs to select vehicle-to-vehicle (V2V) or vehicle-to-edge (V2E) destinations. Experimental evaluation under high-density and highly dynamic conditions demonstrate that the proposed LLM-based scheme outperforms state-of-the-art DRL and Greedy baselines, achieving a 15.3% average reduction in task latency and a 22.1% improvement in energy efficiency over the best DRL baseline, while maintaining a 97.5% task completion rate. Moreover, a fine tuned and quantized deployment reduces inference latency, yielding 1.8$$\times$$ faster decision making at the edge crucial for stringent vehicular deadlines. We discuss remaining challenges, including compute footprint at RSUs, end-to-end latency under bursty loads, and energy aware adaptation, and outline optimization opportunities for real world deployment. Collectively, these results establish LLM-driven offloading as a scalable, accurate, and responsive paradigm for next generation vehicular edge intelligence.