Multi-behavior recommendation systems have seen notable progress by capturing user preferences from diverse behaviors (e.g., view, click, cart). However, existing methods still face the following challenges in accurately and effectively representing multi-behavior semantics: (i) inaccurate intra-behavior feature extraction due to ignoring user-specific interest variations; (ii) limited inter-behavior knowledge transfer caused by behavior distribution gaps. To overcome these issues, we propose a novel model, LAACF (Layer-wise Attention Aggregation based Cascading Fusion), which includes three components: the LAA module for refining user-specific intra-behavior representations via self-attention mechanism; the PEC network for modeling hierarchical inter-behavior dependencies and mitigating behavior gaps; and the KFM module for integrating learned representations and making predictions. Experiments on three benchmark datasets confirm the effectiveness of LAACF compared to existing state-of-the-art methods.