Phase Potential Landscape in LLM Latent Spaces: A Structural Hypothesis for Concept Direction Formation

Fuente: Zenodo

Saved in:

Bibliographic Details
Main Authors:	No-Pattern Engine, Kimminsu, Affiliation: Independent Research
Format:	Recurso digital
Language:	English
Published:	Zenodo 2026
Subjects:	Artificial intelligence Artificial Intelligence/classification phase potential landscape concept direction formation representational tension, compression equilibria latent space geometry cross-model structural similarity (4-(m-Chlorophenylcarbamoyloxy)-2-butynyl)trimethylammonium Chloride steering perturbation attractor stabilization mechanistic interpretability LLM internal geometry training data structure shared generative conditions conditional persistence basin stabilization information geometry (4-(m-Chlorophenylcarbamoyloxy)-2-butynyl)trimethylammonium Chloride/administration & dosage encoding cost gradients logarithmic cost scaling
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Internet