Precision-Guarded Graph-Text Alignment for Universal Chemical Understanding

Large Language Models (LLMs) have demonstrated transformative potential in scientific discovery but frequently suffer from "semantic-structure misalignment"─generating syntactically plausible but chemically invalid structures, or failing to capture precise numerical properties. Existing multimodal adaptations often employ naive projection layers that, under mixed-precision training, lead to feature collapse and the loss of fine-grained topological information. In this work, we propose Deep Graph-Text Alignment (DGTA), a precision-first framework designed to unify structural generation and regression precision. Crucially, we introduce a Stability-Optimized Graph Tokenizer equipped with Float32 Precision Guards and LayerNorm Constraints. Extensive experiments demonstrate DGTA's universality: (1) Quantum Precision: achieving SOTA regression on QM9 (MAE 0.0068); (2) Broad Classification: attaining 79.6% Avg AUC on MoleculeNet; and (3) Generative Robustness: reducing material design error by 59% (MAE 75.53 → 30.83) while achieving 93.16% structural validity on MolQA.