The accurate segmentation of gliomas from multi-sequence magnetic resonance imaging (MRI) remains challenging due to cross-sequence inconsistencies and boundary detection limitations. Although recent neural architectures incorporating convolutional and Transformer-based components have achieved notable advancements, many still exhibit deficiencies in modeling nonlinear inter-sequence interactions and capturing fine-grained boundaries. To address these shortcomings, we propose Extreme Group-Aware Segmentation Network (EGASegNet), a unified 3D segmentation framework that integrates edge-guided attention mechanisms with extreme group-aware transformers for enhanced boundary detection and contextual modeling. The framework introduces three core components that operate synergistically to improve segmentation accuracy. The Adaptive Edge-Guided Module (AEGM) employs boundary-aware spatial-channel joint attention to strengthen anatomically important regions through directional edge focus and adaptive channel recalibration. The Multi-Scale Directional Attention Fusion (MSDAF) module integrates multi-scale directional convolutions with cross-attention mechanisms to reduce cross-sequence misalignment while preserving detailed semantic information across different spatial orientations. The Extreme Group-Aware Attention Former (EGAF) block utilizes partition-based attention with positive-negative value handling to capture long-range dependencies and enhance feature discrimination through entropy-adjusted attention computation. These components are integrated within a hierarchical architecture that ensures coherent semantic refinement across multiple resolution levels, enabling effective fusion of global context and local boundary details for robust tumor segmentation performance. EGASegNet also demonstrates strong generalization performance on the BraTS Lighthouse Challenge testing phase. Our code is available at https://github.com/jlw9999/BraTS2025_GLI .