Move-LLM: Multimodal LLM-Assisted DRL Framework for On-Demand Service Deployment in 6G Vehicular Networks

The Sixth-Generation (6G) vehicular network paradigm demands intelligent solutions for dynamic on-demand service deployment. Yet, current approaches are limited by generalized services for specific traffic events, fixed Roadside Unit (RSU) locations, and insufficient context awareness in unimodal data inputs. This paper proposesMove-LLM, a Multimodal Large Language Model (MLLM)-assisted Deep Reinforcement Learning (DRL) framework for context-awareon-demand service deployment in 6Gvehicular networks. The framework synergizes three core tasks:Perception (P),Recommendation (R), andDeployment (D). For thePtask, raw multimodal (visual, textual) data is fed to a vehicular MLLM (Qwen2-VLM fine-tuned on vehicular events datasets) to enhance its inference via environment perception. For theRtask, the inference result is utilized to recommend customized on-demand services capable of tackling the traffic event. Considering the fixed locations of RSUs, we introduce an Onboard Unit (OBU) cluster formation mechanism, where OBUs form clusters as alternative node options for hosting the recommended services. For theDtask, we develop an improved Deep Deterministic Policy Gradient (DDPG)-based algorithm with the service recommendations as its partial input state, allowing the agent to select the optimal node and resource allocation strategy to deploy on-demand services in a traffic event. Comprehensive simulation results and analysis reveal that our vehicular MLLM significantly improvesPandRtask accuracy by about 18.52% and 4.07%, respectively, in terms of Google BLEU (GLEU) score compared with GPT-4o-mini. Moreover, Move-LLM selects the optimal node and allocates optimal resources, achieving at least 11.76% expected utility under the hybrid (OBU clusters + RSUs) deployment mode compared to the RSU-only mode.