<p>Community detection in complex networks is fundamental for uncovering functionally coherent groups of nodes, with applications in areas such as disease-gene prediction, recommender systems, and social media analysis. However, existing metaheuristic and deep learning–based methods often suffer from high computational cost, sensitivity to parameter tuning, premature convergence, and over-smoothed node embeddings that blur community boundaries, especially in large or noisy networks. This paper proposes GAVAE-GTOLPM, a hybrid framework that couples a label-propagation-based multi-objective Gorilla Troops Optimizer (GTOLPM) with a Graph Attention Variational Autoencoder (GAVAE) for community detection. GTOLPM integrates label propagation, a modularity-based global solution selection strategy, and an adaptive mutation system to accelerate convergence and maintain diverse, well-structured community candidates. GAVAE then learns community-aware node embeddings from an advanced adjacency matrix constructed from GTOLPM solutions and feeds the refined embeddings back to guide the optimizer and avoid local optima. Experiments on six synthetic LFR benchmarks and multiple real-world networks show that GAVAE-GTOLPM consistently achieves the highest or competitive overlapping normalized mutual information (ONMI) and modularity values against ten state-of-the-art baselines. On real-world networks, it improves ONMI over the best competing method by up to 0.048 and modularity by up to 0.111, while maintaining stable performance across runs. These results demonstrate that GAVAE-GTOLPM is a robust and scalable tool for detecting high-quality community structures in complex networks.</p>