Multi-behavior recommendation improves the prediction accuracy of target behaviors (e.g., buy) by leveraging diverse user interactions (e.g., view, collect, cart). This enables more precise modeling of user preferences in personalized recommendation systems. Although existing methods enhance recommendation performance by optimizing cascading behavior relationships, two critical limitations remain. First, current approaches fail to adequately model real-world cascading behavior relationships (e.g., users may skip the “cart” step and proceed directly to buy), resulting in an inability to capture diverse behavioral patterns. Second, while auxiliary behaviors (e.g., view, collect) provide rich supervisory signals, they may also introduce noisy interactions that can negatively impact the prediction of target behaviors. To address these challenges, we propose True-to-Life Behavior Relationship Modeling (TLBRM) for multi-behavior recommendation. TLBRM captures diverse user behavior patterns by modeling realistic shopping scenario cascading behavior relationships (e.g., “view → buy” and “view → collect → buy”). Additionally, we introduce a Prediction-Level Denoising module, which reduces noise in auxiliary behaviors by leveraging supervisory signals from target behaviors, thereby improving recommendation performance. The TLBRM model shows strong performance in predicting target behaviors by effectively utilizing multi-behavior data, as validated on three benchmark datasets: Tmall, Beibei, and Jdata.