Recommender systems have become critical in alleviating information overload, with implicit feedback learning emerging as a dominant approach due to its capacity to capture detailed user behaviors and the simplicity of data collection. However, implicit feedback data is often noisy and biased. Although existing denoising methods have proven effective, they often exacerbate data sparsity by discarding samples or struggling to generalize across different datasets. To address these challenges, we propose a novel implicit feedback recommendation model called the Cross-Model Denoising and Spearman-Based Negative Sample Filling. This model performs data denoising and sample filling by leveraging the collaboration between a primary and an auxiliary model, integrating Kullback-Leibler (KL) divergence and the Spearman rank correlation coefficient. Specifically, noisy samples tend to produce larger prediction discrepancies across models. To capture this effect, we use KL divergence to quantify cross-model prediction consistency, enabling the identification and removal of noisy positive and negative samples while mitigating the bias of single-model judgments. Unlike conventional approaches that rely solely on loss values, which fail to separate noise from informative but hard-to-learn instances, KL divergence effectively reflects discrepancies in predictive distributions under different levels of overfitting, thereby providing a more robust basis for denoising. To mitigate the data sparsity introduced by denoising, we further utilize Spearman rank correlation to select reliable noisy negative samples with higher ranking consistency and refill them into the training set. This not only compensates for data loss but also leverages latent weak signals, enabling the model to better capture users’ implicit interests. Experimental results demonstrate that the proposed model consistently outperforms existing recommendation methods across multiple datasets, enhancing recommendation accuracy.