Beyond the Tokenization Bottleneck: A Conceptual Framework for Efficiency-Plasticity Trade-Offs in ASEAN Large Language Model Adaptation

This paper proposes a conceptual framework to analyze the fundamental trade-offs in adapting Large Language Models (LLMs) to the linguistic diversity of Southeast Asia. Rather than providing new empirical benchmarks, this analytical review introduces Embedding Sparsity Risk (ESR) as a theoretical diagnostic indicator. ESR serves as a conceptual lens for understanding how aggressive vocabulary expansion may lead to under-trained, inactive embedding parameters (often colloquially referred to as "zombie parameters"). By applying this analytical lens, we systematically evaluate three prevailing adaptation pathways: vocabulary expansion, regional full-stack pre-training, and emerging token-free architectures (e.g., MambaByte, Byte Latent Transformer). Drawing on recent multilingual model reports and tokenizer statistics from regional LLM projects, this review clarifies how hardware constraints, licensing thresholds, and linguistic characteristics intersect to shape deployment strategies. To navigate these challenges, we introduce a tiered adaptation roadmap and an open research agenda, shifting the evaluation focus from simple token fertility reduction to optimizing effective parameter utilization and representational robustness in resource-constrained environments.