<p>Vision–language models (VLMs) have shown impressive transferability but still struggle with robustness and generalization when applied to downstream tasks with limited supervision. To address these challenges, we propose a Selective Feature Fusion (SFF) framework that adaptively suppresses noisy visual regions and reinforces task-relevant cross-modal cues through lightweight, learnable gating. Our approach integrates text-guided visual masking and image-aware textual calibration into a unified pipeline, enabling more discriminative and semantically aligned multimodal representations. Comprehensive evaluations across nine widely used benchmarks demonstrate that our method consistently surpasses strong prompt-learning baselines under both few-shot and base-to-novel generalization settings. In particular, under the 8-shot scenario, our approach achieves the best overall accuracy, maintaining a clear margin over representative methods such as CoCoOp and MaPLe. These results highlight not only the robustness of our design but also its effectiveness in capturing cross-modal semantics under data-limited conditions. Further analyses, including ablation studies and qualitative visualizations, confirm that the proposed gating and calibration modules are complementary and play indispensable roles in improving performance. Taken together, this work provides a simple yet powerful strategy for enhancing the adaptability and generalization of VLMs in real-world scenarios.</p>