Recently, multimodal understanding (MMU) and text-to-image generation (T2I) have been integrated into a single autoregressive (AR) architecture, achieving initial unification. However, existing works focus on representation-level studies and overlook potential conflicts in AR architectures' internal information flow during training different tasks. Motivated by this gap, we identify a deeper issue, Task Objective Conflict (TOC), arising from AR architectures' internal information flow, which causes negative transfer and catastrophic forgetting when training MMU and T2I jointly. To address this issue, we proposed UniDecouple, which decouples internal modules for different tasks to construct task-specific optimization subpaths. To implement UniDecouple, we employ a Task-Aware Mixture of Experts (TA-MoE), comprising Hierarchical Expert Routing and Hybrid Expert Collaboration, trained in two stages: first to build task-specific experts, then jointly fine-tuned to balance specialization and overall coordination. Extensive experiments on both understanding and generation benchmarks demonstrate that UniDecouple preserves strong understanding ability while achieving generation quality comparable to state-of-the-art methods, offering a new perspective for unified modeling.