Abstract: Accurate prediction of reservoir parameters is essential for reservoir evaluation. Recent machine learning methods have spurred significant advancements in reservoir prediction; however, limited well logging data and strong reservoir heterogeneity still hinder the accuracy and reliability of such predictions. Addressing these challenges requires methods capable of effectively predicting reservoir parameters under data scarcity and complex reservoir structures. In this study, we propose CAF2, a feature-fusion cross-modal alignment framework guided by large language models (LLMs). CAF2 integrates data augmentation, knowledge distillation, cross-modal alignment, and feature fusion. The data augmentation module employs the RealTabFormer model to generate synthetic datasets that mirror the distribution of real logging data, addressing the challenge of data scarcity. Knowledge distillation extracts essential domain knowledge from LLMs, guiding cross-modal alignment between well logging data and textual knowledge. This alignment mitigates modality gaps via token and sequence alignment, enhancing depth-domain feature representation. Finally, a cross-variable and cross-depth feature fusion strategy exploits both variable information and depth correlations, overcoming the difficulties in accurate reservoir parameter prediction posed by reservoir heterogeneity. Experimental results demonstrate that CAF2 significantly outperforms existing models in predicting tight sandstone reservoir parameters, serving as an effective tool for oil and gas exploration and development.