Guarantees on Dynamical System Distinguishability for LLM Token Generation

• URL: http://arxiv.org/abs/2607.28667v1
• Date: Sat, 25 Jul 2026 04:10:49 GMT
• Title: Guarantees on Dynamical System Distinguishability for LLM Token Generation
• Abstract summary: We formalize the classification task as a binary hypothesis test between two linear DSs.<n>We show that the total variation distance between the stationary marginal distributions of the two DSs can be arbitrarily small.<n>We also characterize cross-embedding generalization by introducing an approximate intertwining condition between embedding models.
• Score: 3.6209977238182276
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent work has shown that classifying large language models (LLMs)' responses can be distinguished by modeling token embeddings as trajectories of a black-box dynamical system (DS) and comparing prediction residuals of two DSs. Despite the empirical success of this dynamical approach, a theoretical understanding of why it works, how well it scales as a function of the token sequence, and when it transfers across embedding models remains lacking. We address these questions by formalizing the classification task as a binary hypothesis test between two stochastic linear DSs. We show that the total variation distance between the stationary marginal distributions of the two DSs can be arbitrarily small even when the dynamics differ substantially, which provides a fundamental accuracy floor for any classifier that ignores token dynamics. We then show that the misclassification probability of DS-based classification decays exponentially in the sequence length $L$, with the decay governed by a dynamical discriminability quantity $δ^2$ that captures the spectral distance between the two DSs. We also characterize cross-embedding generalization by introducing an approximate intertwining condition between embedding models and establishing a lower bound on the transferable discriminability in terms of the intertwining map's smallest singular value. Together, these results explain the empirical performance of DS-based classification and motivate further investigation into using DS theory to analyze AI systems, in contrast to the more common approach of using AI to model dynamical systems.

Related papers

• Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts [60.95873409673775]
  identifiability in continuous-time latent differential equation models remains largely open.<n>We study additive-noise latent SDEs observed through an unknown nonlinear diffeomorphism.<n>We propose a two-stage estimator for latent disentanglement and optional graph recovery.
  arXiv  Detail & Related papers  (2026-06-26T16:18:28Z)
• Effective Covariance Dynamics in Solvable High-Dimensional GANs [2.639628765667673]
  We study a solvable high-dimensional model of generative adversarial network (GAN) training in which a linear generator learns a low-dimensional subspace from data.<n>We prove that the microscopic training process converges, in the high-dimensional limit, to deterministic ordinary differential equations governed by this effective covariance.
  arXiv  Detail & Related papers  (2026-06-25T16:33:54Z)
• Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine [60.669081685261965]
  Diffusion models generate high-dimensional data with remarkable quality.<n>How their training efficiently learns the score function remains theoretically unexplained.<n>We instantiate this principle as Score-induced Latent Diffusion (SiLD)
  arXiv  Detail & Related papers  (2026-05-16T16:51:10Z)
• Towards Fast Coarse-graining and Equation Discovery with Foundation Inference Models [6.403678133359229]
  latent dynamics in high-dimensional recordings are often characterized by a much smaller set of effective variables.<n>Most machine learning approaches tackle these tasks jointly by training autoencoders together with models that enforce dynamical consistency.<n>We propose to decouple the two problems by leveraging the recently introduced Foundation Inference Models (FIMs)<n>A proof of concept on a double-well system with semicircle diffusion, embedded into synthetic video data, illustrates the potential of this approach for fast and reusable coarse-graining pipelines.
  arXiv  Detail & Related papers  (2025-10-14T15:17:23Z)
• Understanding Catastrophic Interference: On the Identifibility of Latent Representations [67.05452287233122]
  Catastrophic interference, also known as catastrophic forgetting, is a fundamental challenge in machine learning.<n>We propose a novel theoretical framework that formulates catastrophic interference as an identification problem.<n>Our approach provides both theoretical guarantees and practical performance improvements across both synthetic and benchmark datasets.
  arXiv  Detail & Related papers  (2025-09-27T00:53:32Z)
• Hierarchical Stochastic Differential Equation Models for Latent Manifold Learning in Neural Time Series [0.0]
  We propose a novel hierarchical differential equation (SDE) model that balances computational efficiency and interpretability.<n>We derive training and inference procedures and show that the computational cost of inference scales linearly with the length of the observation data.
  arXiv  Detail & Related papers  (2025-07-29T06:51:58Z)
• Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning [80.44084021062105]
  We propose a novel latent partial causal model for multimodal data, featuring two latent coupled variables, connected by an undirected edge, to represent the transfer of knowledge across modalities.<n>Under specific statistical assumptions, we establish an identifiability result, demonstrating that representations learned by multimodal contrastive learning correspond to the latent coupled variables up to a trivial transformation.<n>Experiments on a pre-trained CLIP model embodies disentangled representations, enabling few-shot learning and improving domain generalization across diverse real-world datasets.
  arXiv  Detail & Related papers  (2024-02-09T07:18:06Z)
• Learning Stochastic Dynamical Systems as an Implicit Regularization with Graph Neural Networks [5.374505641331418]
  Gumbel graph networks are proposed to learn high-dimensional time series. We show that S-GGNs exhibit superior convergence, robustness, and generalization, compared with state-of-the-arts networks.
  arXiv  Detail & Related papers  (2023-07-12T11:38:34Z)
• A Geometric Perspective on Diffusion Models [57.27857591493788]
  We inspect the ODE-based sampling of a popular variance-exploding SDE. We establish a theoretical relationship between the optimal ODE-based sampling and the classic mean-shift (mode-seeking) algorithm.
  arXiv  Detail & Related papers  (2023-05-31T15:33:16Z)
• Identifiability and Asymptotics in Learning Homogeneous Linear ODE Systems from Discrete Observations [114.17826109037048]
  Ordinary Differential Equations (ODEs) have recently gained a lot of attention in machine learning. theoretical aspects, e.g., identifiability and properties of statistical estimation are still obscure. This paper derives a sufficient condition for the identifiability of homogeneous linear ODE systems from a sequence of equally-spaced error-free observations sampled from a single trajectory.
  arXiv  Detail & Related papers  (2022-10-12T06:46:38Z)
• Linearization and Identification of Multiple-Attractors Dynamical System through Laplacian Eigenmaps [8.161497377142584]
  We propose a Graph-based spectral clustering method that takes advantage of a velocity-augmented kernel to connect data-points belonging to the same dynamics. We prove that there always exist a set of 2-dimensional embedding spaces in which the sub-dynamics are linear, and n-dimensional embedding where they are quasi-linear. We learn a diffeomorphism from the Laplacian embedding space to the original space and show that the Laplacian embedding leads to good reconstruction accuracy and a faster training time.
  arXiv  Detail & Related papers  (2022-02-18T12:43:25Z)
• BCD Nets: Scalable Variational Approaches for Bayesian Causal Discovery [97.79015388276483]
  A structural equation model (SEM) is an effective framework to reason over causal relationships represented via a directed acyclic graph (DAG) Recent advances enabled effective maximum-likelihood point estimation of DAGs from observational data. We propose BCD Nets, a variational framework for estimating a distribution over DAGs characterizing a linear-Gaussian SEM.
  arXiv  Detail & Related papers  (2021-12-06T03:35:21Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.