Abstract:Multi-agent reinforcement learning (MARL) holds broad application prospects in fields such as collective intelligence modeling, complex collaboration, and adversarial decision-making. However, existing structured communication methods based on graph neural network (GNN) are prone to over-smoothing in deep propagation and the communication topology is difficult to update adaptively during training, which limits communication efficiency and generalization performance. To this end, this study proposes RESC-LSC, a unified communication framework integrating residual-enhancement and dynamic topology learning based on the classic layered communication architecture LSC (learning structured communication). This framework consists of two key components: the DRGL module adaptively reconstructs the communication graph via differentiable semantic similarity modeling, and the REGC module incorporates initial residual injection and identity mapping mechanisms based on the GCNII structure to enhance the stability of deep feature propagation and mitigate over-smoothing. The two modules act synergistically at the two levels of structural learning and feature propagation, thus improving the expressiveness and task adaptability of the communication graph. Experimental results demonstrate that compared with the original LSC and its two ablated variants, the RESC-LSC framework achieves stable and consistent performance improvements in key indicators including average reward, communication efficiency, and structural stability, indicating the effectiveness and scalability of the proposed method for complex multi-agent communication modeling.