Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation | ArxivCSExplorer