Efficiency-Fidelity Trade-offs in Legal Document Generation: Evaluating Optimization Strategies for Vietnamese Small Language Models with Normative Content Preservation Analysis

Deploying language models for legal document processing in emerging jurisdictions confronts a critical question: do computational efficiency gains compromise normative legal content? This study systematically evaluates optimization strategies for Vietnamese Small Language Models across legal summarization and title generation tasks. We compare three model families (ViT5, BARTpho-word, BARTpho-syllable) under four training configurations (Standard, Dynamic Padding, Layer Freezing, Hybrid) using 48,320 Vietnamese legal documents. Beyond standard metrics, we introduce a normative content preservation protocol assessing retention of deontic markers and legal holdings through expert annotation. Results demonstrate that Hybrid optimization achieves 87.4% carbon emission reduction and 2.01\((\times)\) speedup while maintaining semantic fidelity statistically indistinguishable from baseline. Critically, optimized large models preserve 88% of deontic content versus 91% for unoptimized versions. Optimized large models outperform inherently smaller variants on both efficiency and accuracy (BARTpho-syllable Hybrid: 0.8521 vs Base: 0.8378). Ablation analysis reveals 40% layer freezing as optimal. Error analysis shows optimization increases argumentative simplification but not hallucination (3% constant), indicating safer failure modes for legal applications. These findings establish that aggressive optimization preserves legally salient content, enabling sustainable deployment of legal AI in resource-constrained civil law jurisdictions.