Semantic Bundle AI: Anchor Design and Stability Framework — Stable Semantic Regions, Domain Discriminability, and a Four-Axis Evaluation Framework for Semantic Stability
Year: 2026
Type: preprint
Abstract: Semantic Bundle AI introduces stable anchor-based coordinates as a complementarysemantic management layer for Large Language Models [4,5]. A key open question from priorwork concerns anchor design: what constitutes an effective anchor, and how should anchorsbe selected?This paper addresses that question through two experiments. First, we demonstrate thatanchors constructed as stable semantic regions — averaged embeddings of meaning-neighborclusters rather than single words — achieve substantially higher inter-model stability thanword-level anchors (Spearman r = 0.704 vs. 0.101, a 7× improvement). Second, we identify afundamental tradeoff between cluster stability and domain discriminability: generic anchorsexcel at intra-cluster cohesion (variance ratio 0.047–0.067), while domain-specific conceptvectors achieve 4.4× higher inter-domain separation (0.739 vs. 0.166). We propose a four-axisevaluation framework for semantic stability — inter-model consistency, temporal consistency,cross-cultural robustness, and perturbation resistance — and provide initial validation ontwo axes. These results establish concrete design principles for anchor selection and positionsemantic stability as an empirically tractable research programCorrected a metric conflation in the Discussion (H-102/ranking consistency); core results unchanged.
Author: saitou, Makoto
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162697962