<p>Vision-language (VL) tracking leverages natural language descriptions to provide richer semantic guidance for target localization in video sequences. However, prevailing methods exhibit notable limitations in spatial detail modeling, cross-modal semantic correlation, and feature discriminability. Specifically, they often inadequately fuse fine-grained local geometric structures with linguistic cues, while deep features are susceptible to redundancy and collapse, thereby constraining tracking robustness. To address these issues, this paper proposes a novel VL tracking model, termed CET-Track, which systematically enhances performance through the synergistic optimization of the Transformer decoder. Our contributions are threefold: First, we design a WindowAttention-sparse Module to strengthen local spatial modeling with reduced computational overhead. Second, we introduce a Relation-Aware Attention Module that explicitly incorporates learnable semantic relationships into the attention computation, fostering profound language-vision interaction. Finally, we devise a Contra Norm Module that integrates contrastive learning into the feature normalization process, effectively mitigating feature redundancy and collapse while boosting discriminative power. Extensive experiments on four mainstream VL tracking benchmarks—TNL2K, LaSOT, LaSOText, and OTB99-Lang—demonstrate that CET-Track surpasses state-of-the-art methods across multiple evaluation metrics. It achieves an advantageous balance between tracking precision and model efficiency, offering an effective solution for robust vision-language tracking in complex scenarios.</p>