The rapid advancement of smart cities and the Internet of Things (IoT) has led to an explosive growth in connected devices, driving an urgent demand for communication networks with low latency and energy consumption. In multi-cell environments, conventional Orthogonal Multiple Access (OMA) schemes struggle to accommodate real-time communication for massive devices due to limited bandwidth resources. While computation offloading techniques have shown robustness in handling standardized tasks, most existing approaches overlook user-specific preferences and lack interactive capability. To address these limitations, we propose LAM-MD , a M ulti-agent D eep reinforcement learning (MADRL) framework integrated with a Large Action Model (LAM) . The framework jointly optimizes latency and energy consumption in multi-cell Non-Orthogonal Multiple Access (NOMA) networks, while actively integrating personalized user demands. By interpreting natural language inputs, the LAM acts as a user-demand translator to generate tailored offloading strategies. These strategies are subsequently processed by our AdaMD algorithm, which computes user-specific optimal policies. Furthermore, by incorporating an Adaptive Classification-based Experience Replay (ACER) mechanism, our approach accelerates convergence, enhances the MADRL architecture, and significantly boosts the effectiveness of LAM-driven decision-making. Extensive experiments validate that the proposed framework outperforms existing methods in both delay and energy optimization, while demonstrating strong real-time performance and scalability. Numerical results further confirm that our model achieves a substantial reduction in overall cost, encompassing both energy consumption and latency, compared to state-of-the-art baselines.
He et al. (Wed,) studied this question.