Software reliability has continued to be one of the most acute problems in contemporary software engineering, especially as codebase complexity grows, release cycles speed up and the workforce grows more heterogeneous. This paper presents an adaptive optimization framework multimodal software defect prediction design based on reinforcement learning that can dynamically control the combination of heterogeneous data modalities. However, as opposed to standard fusion strategies, the reinforcement learning agent will constantly modify the modality contributions, as it gets feedback on their performance measurements, such as F1-score and false negative rate.  The method can improve accuracy, robustness, as well as establishing a base towards scalable, interpretable, and self-optimizing defect prediction systems in industrial software analytics.