Cross-Modal Adaptive Reasoning for Long-Tailed Visual-Linguistic Understanding

Vision-language models often struggle in practical applications because real-world data follow long-tailed distributions, leaving many rare visual concepts poorly represented. This imbalance leads to biased feature learning and weak semantic alignment for infrequent categories. To address these challenges, we introduce Cross-Modal Adaptive Reasoning for Long-Tailed Visual-Linguistic Understanding (CARL-VU), a framework that strengthens model robustness and generalization, particularly for rare concepts and complex semantic relations. CARL-VU combines a transformer-based encoder–decoder design with a semantic-guided expert routing mechanism that dynamically selects specialized experts based on input content. It further incorporates contrastive distillation to enhance the distinctiveness of tail-class features and adaptive feature augmentation to enrich data diversity. Through a two-stage training scheme, the model learns to handle a wide range of visual–linguistic inputs more effectively. Experiments on long-tailed benchmarks demonstrate clear improvements over existing approaches, and ablation analyses verify the complementary contributions of each component in alleviating long-tail issues.