This paper presents a hybrid deep learning framework that integrates a Large Language Model (LLM) with a Transformer architecture for stock price prediction. The authors address a theoretical gap in prior work that combines textual and numerical data empirically without formalizing their interaction. They conceptualize a prompt-based LLM as a mathematically defined signal generator that extracts directional market sentiment and an associated confidence score from financial news. These signals are dynamically fused with historical price features through a noise-robust gating mechanism, enabling the Transformer to adaptively weigh semantic and quantitative information. The proposed approach is evaluated on stock price forecasting tasks, demonstrating a 5.28% reduction in Root Mean Squared Error (RMSE) over a Vanilla Transformer baseline, with statistical significance (p=0.003). Ablation studies confirm the critical role of the gating mechanism and LLM signals, while robustness analysis shows the model's stability under noisy conditions. Additionally, the confidence-weighted attention mechanism provides interpretability by highlighting influential news and price features. The findings support a paradigm shift from empirical observation to formalized modeling of LLM-Transformer interactions, paving the way toward explainable, noise-resilient, and semantically enriched financial forecasting systems.