From the perspective of industrial information integration engineering (IIIE), the Industrial Internet-of-Things (IIoT) serves as a unified framework that integrates cloud, edge, and manufacturing resources through cloud–edge–device collaboration, enabling highly flexible and collaborative production processes. Collaborative task scheduling in IIoT refers to assigning manufacturing and computational tasks to heterogeneous resources to minimize the overall task makespan and energy consumption. However, the presence of complex task dependencies and the heterogeneity of resource configurations make the scheduling problem highly challenging. To address this, we conduct a comprehensive evaluation of seven evolutionary algorithms (EAs) and seven deep reinforcement learning (DRL) methods across three representative IIoT scheduling scenarios: manufacturing task scheduling (MTS), computational task scheduling (CTS), and hybrid task scheduling (HTS). To investigate the effect of algorithm design, we propose two types of algorithm formulations: explicit formulation (EF), where the algorithm outputs correspond directly to decision variables, and implicit formulation (IF), where outputs represent heuristic factors guiding task assignment. For each scenario, we construct scheduling instances of three scales and evaluate all 14 methods under both formulations. The results demonstrate that EAs offer more stable performance, while DRLs exhibit stronger generalization and faster inference, especially in large-scale or dynamic scenarios. Moreover, the implicit formulation often leads to better solution quality across both algorithm classes. These findings provide valuable insights for algorithm selection and design in IIoT environments and highlight the importance of formulation strategies in influencing optimization outcomes.