Large Language Models (LLMs) have achieved remarkable performance across a wide range of natural language processing tasks; however, their underlying semantic representations remain fundamentally relative and context-dependent. This reliance on relative representations imposes structural limitations on long-term semantic stability, reproducibility, and controllable semantic editing, with semantic drift during retraining or fine-tuning being a representative symptom.This paper introduces Meaning Bundle AI, a novel framework that stabilizes semantic representations by anchoring them to a shared coordinate system based on stable semantic anchors. Importantly, Meaning Bundle AI is not intended to replace Transformer-based architectures or relative embedding methods used in LLMs. Rather, it is designed as a complementary layer that presupposes existing models and augments them with long-term referential stability, reproducibility, and efficient semantic management.Instead of storing semantic information holistically within model parameters, Meaning Bundle AI adopts a referential reconstruction approach, in which semantic representations are reconstituted on demand from a stable coordinate system and a compact set of parameters. This design preserves compatibility with existing embedding-based architectures while enabling substantial reductions in parameter usage, computational overhead, and retraining frequency.To prioritize stability and verifiability, this paper intentionally presents a minimal mathematical formulation, omitting non-essential components such as complex relation tensors and nonlinear elastic extensions. The proposed framework does not negate existing stabilization techniques but rather provides a unifying geometric perspective that integrates them at a higher conceptual level. We further outline proof-of-concept (PoC) designs for evaluating temporal stability, reproducibility, and efficiency.Detailed mathematical formulations, algorithmic designs, and extended discussions are provided in the Japanese version of this paper.
makoto saitou (Tue,) studied this question.