From Attention to Dynamics: A Structural Analysis of Feedback, Coherence, and State in Modern AI Architectures

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteur principal: Ryder, John F.
Format: Recurso digital
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901539327574016
author Ryder, John F.
author_facet Ryder, John F.
contents <p>This paper argues that current AI interpretability methods are structurally limited by their reliance on scalar probes such as attention weights and activation magnitudes. These tools are incapable of detecting directional, phase-sensitive, and temporally extended coupling dynamics that may govern system behaviour.</p> <p>We introduce the concept of an observable kernel mismatch: systems may exhibit real internal structure that lies outside the detectable class of the measurement tools applied. Drawing on both dynamical systems theory and empirical precedent from astrophysical observation, we show that state space models (SSMs) provide a natural substrate for such dynamics through continuous-time latent state evolution governed by matrix coupling.</p> <p>We propose a set of architectural constraints for dynamically coherent systems and outline testable predictions using dynamical measurement tools such as transfer entropy, Granger causality, and topological analysis.</p> <p>This work reframes AI systems as dynamical systems requiring appropriate measurement frameworks, rather than static function approximators fully characterisable by scalar analysis.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19659575
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle From Attention to Dynamics: A Structural Analysis of Feedback, Coherence, and State in Modern AI Architectures
Ryder, John F.
AI interpretability; dynamical systems; state space models; transformers; neural architectures; observable kernel; transfer entropy; Granger causality; topological data analysis; machine learning theory
<p>This paper argues that current AI interpretability methods are structurally limited by their reliance on scalar probes such as attention weights and activation magnitudes. These tools are incapable of detecting directional, phase-sensitive, and temporally extended coupling dynamics that may govern system behaviour.</p> <p>We introduce the concept of an observable kernel mismatch: systems may exhibit real internal structure that lies outside the detectable class of the measurement tools applied. Drawing on both dynamical systems theory and empirical precedent from astrophysical observation, we show that state space models (SSMs) provide a natural substrate for such dynamics through continuous-time latent state evolution governed by matrix coupling.</p> <p>We propose a set of architectural constraints for dynamically coherent systems and outline testable predictions using dynamical measurement tools such as transfer entropy, Granger causality, and topological analysis.</p> <p>This work reframes AI systems as dynamical systems requiring appropriate measurement frameworks, rather than static function approximators fully characterisable by scalar analysis.</p>
title From Attention to Dynamics: A Structural Analysis of Feedback, Coherence, and State in Modern AI Architectures
topic AI interpretability; dynamical systems; state space models; transformers; neural architectures; observable kernel; transfer entropy; Granger causality; topological data analysis; machine learning theory
url https://doi.org/10.5281/zenodo.19659575