In recent years, transformer-based human-object interaction (HOI) detection methods have gained widespread attention for their excellent relational modeling capabilities. However, existing methods generally adopt static fusion strategies in the feature enhancement stage, which makes it difficult to fully explore the fine-grained interaction information in multi-relational contexts, limiting the inference ability and generalization performance of the model in complex scenarios. To this end, we propose a cross-attention augmentation module (CAAM) to enhance the selectivity and expressiveness of semantic fusion by introducing a task-aware cross-attention mechanism to establish dynamic interaction paths between task features and multi-relational contexts. The module can be flexibly embedded into existing multi-branch HOI detection frameworks without introducing additional supervision. Experiments on the V-COCO dataset show that the proposed method can stably improve interaction detection accuracy, demonstrating its potential for application in complex interaction understanding tasks.