Large Language Models (LLMs) demonstrate strong generative capabilities, yet remain vulnerable to semantic drift, hallucination, and instability in long-horizon reasoning. This study proposes Semantic Drafting (SD), a framework that redefines prompting as a structured semantic control system. SD organizes inference through hierarchical layers—L4 (objective), L3 (context), L2 (constraints), and L1 (problem)—which collectively bound the reasoning space. In addition, this study introduces a structural interpretation of hallucination based on Semantic Flexibility (SF). Rather than treating hallucination as random error, we classify it as a set of failure modes arising from imbalances in semantic integration. Through controlled experiments with structured and ablated conditions, we show that removing SD does not merely degrade performance but shifts output distributions toward distinct failure regimes. These findings suggest that LLM instability is fundamentally a structural phenomenon governed by constraint design and semantic flexibility.
卓馬 吉田 (Sun,) studied this question.