K-Field Routing: Cross-Layer Causal Signals for Multi-Hop Reasoning in Transformers
Fuente:
Zenodo
Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Pubblicazione: |
Zenodo
2026
|
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866901065742417920 |
|---|---|
| author | Li, Y.Y.N. |
| author_facet | Li, Y.Y.N. |
| contents | <p>We propose a theoretical framework in which neural network optimization dynamics can be characterized by an effective Lorentzian spacetime structure in the reduced $(K,\sigma)$ state space—analogous to the (1+1)-dimensional geometry of special relativity. The Hessian of a reference Lyapunov potential exhibits signature $\Sig(\Hessian)=(1,1)$, where the information flow ratio $K$ plays the role of 'time' and entropy $\sigma$ of 'space'. This geometric structure, combined with stability requirements, theoretically constrains the space of optimal neural architectures.</p> <p>We present \textbf{K=1 Chronogeometrodynamics}, a theoretical framework deriving optimal neural structures from information-geometric first principles. Our contributions are threefold: (1) a \textbf{uniqueness theorem} proving that Lorentzian signature $\Sig(\Hessian)=(1,1)$ with stability constraints forces a unique optimal structure $\Jmat_{\Hessian} = \alpheff \Hessian^{-1} \Jmat$, (2) definition of the K-metric ($K = \dd{\Phi}/H$) as a training diagnostic, and (3) experimental validation providing empirical support for dissipative dynamics with negative average Lyapunov drift.</p> <p>Experiments show K-metric convergence from high initial values toward task-dependent equilibria, with negative average drift supporting the predicted dissipative behavior. This work represents a step from \emph{trial-and-error architecture design} toward \emph{geometry-guided principles}.</p> <p>\textbf{Keywords}: Neural architecture, information geometry, port-Hamiltonian systems, Lorentzian geometry, optimal control, deep learning theory</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18970383 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | K-Field Routing: Cross-Layer Causal Signals for Multi-Hop Reasoning in Transformers Li, Y.Y.N. <p>We propose a theoretical framework in which neural network optimization dynamics can be characterized by an effective Lorentzian spacetime structure in the reduced $(K,\sigma)$ state space—analogous to the (1+1)-dimensional geometry of special relativity. The Hessian of a reference Lyapunov potential exhibits signature $\Sig(\Hessian)=(1,1)$, where the information flow ratio $K$ plays the role of 'time' and entropy $\sigma$ of 'space'. This geometric structure, combined with stability requirements, theoretically constrains the space of optimal neural architectures.</p> <p>We present \textbf{K=1 Chronogeometrodynamics}, a theoretical framework deriving optimal neural structures from information-geometric first principles. Our contributions are threefold: (1) a \textbf{uniqueness theorem} proving that Lorentzian signature $\Sig(\Hessian)=(1,1)$ with stability constraints forces a unique optimal structure $\Jmat_{\Hessian} = \alpheff \Hessian^{-1} \Jmat$, (2) definition of the K-metric ($K = \dd{\Phi}/H$) as a training diagnostic, and (3) experimental validation providing empirical support for dissipative dynamics with negative average Lyapunov drift.</p> <p>Experiments show K-metric convergence from high initial values toward task-dependent equilibria, with negative average drift supporting the predicted dissipative behavior. This work represents a step from \emph{trial-and-error architecture design} toward \emph{geometry-guided principles}.</p> <p>\textbf{Keywords}: Neural architecture, information geometry, port-Hamiltonian systems, Lorentzian geometry, optimal control, deep learning theory</p> |
| title | K-Field Routing: Cross-Layer Causal Signals for Multi-Hop Reasoning in Transformers |
| url | https://doi.org/10.5281/zenodo.18970383 |