<p>Multimodal recommender systems enhance recommendation accuracy by integrating multi-source information such as text and vision, but existing methods face the challenges of modal noise pollution and rigid fusion mechanism. In this study, we propose a multimodal recommendation algorithm based on hard-negative sample mining and adaptive fusion (HNAF), which achieves accurate modeling of user preferences through feature purification, dynamic modality fusion, and hard-case augmented learning. The HNAF consists of three core mechanisms: a user perception purifier, which is guided by behavioral ID embedding, and filters out the redundant noise in the modality features through the attentional mechanism to retain the semantically relevant information; Adaptive early and late fusion module, which combines the bottom-level feature interaction of early fusion with the high-level semantic integration of late fusion, and dynamically balances the contributions of both through learnable parameters to capture cross-modal multilevel associations; Hard negative sample mining is introduced into contrast learning, which filters the hardest cases to the positive samples from global negative samples, and strengthens the feature discriminative power through the loss of hybrid weighting to improve the model’s ability to distinguish similar items. Experiments on three real datasets verify the effectiveness of the proposed model, with all metrics across the three datasets improving by more than 3.5%.</p>