External Navigation of LLM Latent Space via Multi-Domain Probing

We propose a method for externally navigating the latent space of large language models(LLMs) without access to internal representations. Our approach uses multi-domain probing:sending semantically identical queries from multiple disciplinary perspectives and extracting coordinateshigh-frequency anchor words that appear across domains. We demonstratethat (1) stable places exist in LLM output space, evidenced by Jaccard similarity of 50.8%across repeated queries at temperature 0; (2) cross-architecture invariants emerge acrossClaude, GPT, and Llama families, with core coordinates such as gravity achieving 100%frequency; (3) holographic probingsimultaneously asking WHERE, WHAT, HOW, WHY,WHERE-FROM, and WHERE-TO within a single promptsystematically elevates abstraction from components to systemic architecture, producing 1928 unique coordinates perconcept invisible to other formats; (4) the method produces results 35× above formal random baseline (z > 6, p < 0.001 for 17/18 experiments); and (5) cascade prediction achievesveriable structural extrapolation with self-calibrating accuracy. We introduce a taxonomyof ve question patterns and present quantitative results from a platform implementing themethod across 16 models and 3 architecture families.