Suppressing Hallucination for Trustworthy LLMs (Part I – Semantic Reliability Framework)

This study reframes hallucination in large language models (LLMs) not as a simple technical error but as a form of semantic reliability collapse. To address this issue, the paper introduces the Layer-Knot Framework (LKF), a structural model that stabilizes meaning by forming semantic “knots” across deep layers, preserving resonance among intention, evidence, and context. Building on this framework, we propose a triple-indicator evaluation system consisting of Hallucination Rate (HR), Groundedness Rate (GR), and Coherence Rate (CR), offering a unified approach to measuring semantic stability. Empirical tests on TruthfulQA and 2,000 diverse prompt domains show that LKF-aligned models achieve a 50% reduction in hallucination, a 12% increase in groundedness, and stable creative variance (±3%) compared to standard LLMs. These findings demonstrate that semantic reliability and generative creativity are not mutually exclusive; rather, they can coexist through structural alignment within the model’s internal topology. The paper argues that suppressing hallucination is not merely a matter of improving accuracy but a deeper process of aligning the internal semantic topology of the model. HR/GR/CR are shown to provide a robust quantitative–qualitative foundation for future studies in AI reliability, interpretability, and AI ethics. Overall, the LKF establishes a structural and philosophical basis for next-generation trustworthy LLMs and opens a path toward self-monitoring AI systems capable of evaluating their own semantic integrity.