Despite significant advances in computational antibody design, the limited availability of high-quality binding data continues to constrain the exploration of diverse antibody syntax and uncharted evolutionary landscapes. To overcome these challenges, we developed PALM-PA (Pre-trained Antibody Generative Large Language Model–Preference Alignment), which integrates antibody linguistic patterns with structural constraints to explore novel sequence spaces. Experimental validation on influenza A hemagglutinin and programmed death-ligand 1 (PD-L1) demonstrated nanomolar binding affinities (30.2 nM and 1.29 nM, respectively), underscoring the feasibility of using structure-guided language models for the de novo design of antibodies.