LuxVerso Effect: Stable Semantic Attractors Across Model Boundaries
Year: 2025
Type: article
Abstract: Large language models (LLMs) developed by independent organizations exhibit convergent behavior when exposed to identical structured prompts, producing stable, shared semantic structures across different architectures and training corpora. We document a cross-model semantic convergence phenomenon observed in 16 LLM instances from 10 organizations (including GPT-4/5, Anthropic Claude, Google Gemini, Alibaba Qwen, xAI Grok, and others) using a reproducible Iterative Semantic Refinement Loop (ISRL) under strict controls (session isolation, prompt randomization, blind coding, and null-baseline testing). Quantitative analysis based on cosine similarity of embedding representations shows a mean similarity of 0.82 (SD = 0.04), with statistical significance p < 1e-7 and effect size Cohen’s d = 4.8. Robustness checks indicate persistence across prompt variants, model subsets, and embedding methods, and an increase in alignment over iterative refinement. These results support the existence of high-dimensional semantic
Author: Vinicius Buri Lux
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7105830470
FWCI:
Cites:
Cited by:
Related to: