Abstract: Rolling monthly (1-30 days) precipitation prediction falls within the sub-seasonal to seasonal (S2S) timescale, a range often characterized by limited predictability. A key challenge lies in detecting and correcting systematic biases embedded in dynamical model outputs, which are strongly influenced by internal variability and noise. To address this challenge, we develop a deep learning bias-correction framework based on a hybrid CNN-Transformer architecture (H2Former). The framework leverages the strengths of Convolutional Neural Networks (CNNs) in capturing local fine-scale features, while utilizing Transformers to model long-range dependencies associated with large-scale circulation patterns. To comprehensively evaluate its performance, we applied the framework to forecasts from NCEP-CFSv2 and conducted systematic bias-correction experiments, input-factor ablation studies, and architectural comparisons, using the Yangtze River Basin as a representative testbed. Results on an independent 2022-2024 test set show that forecast accuracy across the Yangtze River Basin improved from 62.1% to 64.8%, with gains of up to 8.2 percentage points in the topographically complex lower Jinsha River region. The ablation experiments further indicate that the dynamical model’s raw precipitation forecasts are the dominant source of skill, highlighting the role of deep learning in identifying and correcting spatiotemporally structured systematic biases. Comparative analyses also confirm the superiority of the hybrid H2Former design. Overall, this study demonstrates the potential of deep learning as an intelligent post-processor for dynamical climate models to enhance S2S precipitation trend prediction skill. The proposed framework provides methodological support for developing refined “qualitative-then-quantitative” climate prediction services and offers substantial application value for water resource management and disaster risk reduction in the Yangtze River Basin.