On physical implementability of reinforcement-learning-driven reasoning

Eiji Konishi

Home/Archive/arXiv

arXiv · 2610.14131

On physical implementability of reinforcement-learning-driven reasoning

Eiji Konishi

View original source ↗Read source document ↗Export metadata ↓

Abstract

In the machine learning community, the reinforcement-learning-driven reasoning process has been intensively studied since around 2025. In this article, we present a negative result on the hardware implementation of this reinforcement-learning-driven reasoning process in a physical neural network model in the presence of a reward external potential energy for a given reward. In this hardware implementation, the existence of a chain of thought in dynamics contradicts the Lyapunov stability of the system.

Explore related subjects

cond-mat.dis-nn cs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Eiji Konishi. 2026-10-09. On physical implementability of reinforcement-learning-driven reasoning. https://arxiv.org/abs/2610.14131

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Two-stage Slowdown of Vacancy-mediated Diffusion in a Lattice Gas Model with Quenched Disorder and Spins

We present a lattice gas model with random interactions quenched in space and spins quenched to the particles. The vacancy-mediated diffusion is studied in equilibrium. A two-stage slowdown is witnessed in both the mean squared displacement and the spatial overlaps. We demonstrate that the plateaus are due to anti-correlation at the corresponding length scale. The anti-correlation diminishes quickly with increasing lag time. A broad and yet finite distribution of the survival function of vacancies suggests their normal diffusion at large time limit. Our model can serve as a new playground for understanding hierarchical dynamics in glassy-like dynamical systems.

cond-mat.dis-nn↗

arXiv2026

Resonance Proliferation Across Localization Transitions

Models of many-body localization (MBL) exhibit slow numerical drifts towards delocalization with increasing system size, for which no satisfactory theory exists. Numerical results indicate that these drifts are driven by the proliferation of many-body resonances at intermediate disorder strengths. We develop a statistical method to predict the distribution of resonance oscillation frequencies, accounting for how resonances formed at larger frequency scales affect the subsequent formation of resonances at lower frequencies. Working within the statistical Jacobi approximation (SJA), we derive a flow equation for a power-law exponent $θ(w)$ characterizing the density of resonances at frequency scale $w$. A localized phase is described by a line of fixed points with $θ(w)>0$, while an instability of the flow signals resonance proliferation and the onset of thermalization. The predicted $θ(w)$ agrees with numerical results for the Anderson model on random regular graphs and the Lévy--Rosenzweig--Porter random matrix ensemble, both of which host resonance-driven delocalization transitions. We further connect the flow to eigenstate properties such as the participation ratio and to dynamical observables such as the return probability. The predicted $θ(w)$ also agrees with numerical measurements in real-space models of MBL at intermediate disorder strengths, providing a step towards explaining the finite-size drifts observed in MBL.

cond-mat.dis-nn↗

arXiv2026

Moments of Random Permanents and Hafnians: A Replica Approach

Using contour-integral representations of the permanent and of the Hafnian, we turn the computation of their moments for random matrices into a replica calculation, analogous to that of the moments of the partition function of a mean-field spin glass. For the $2k$-th moment of $N\times N$ matrices with independent entries, the dominant saddle point is replica symmetric for $k>2$, and we obtain in closed form the exponential growth rate, the prefactor and the first $1/N$ correction, in agreement with exact enumerations. The growth rate depends only on whether the entries are real or complex; higher cumulants affect only the prefactor and the $1/N$ term. At $k=2$ the replica-symmetric solution becomes unstable, separating a power-law bulk of the distribution from a large-deviation tail, which we compute. The results apply to boson sampling with single photons and with squeezed light.

cond-mat.dis-nn↗