Stochastic Dynamics of Generative AI: From Autoregressive Models to Interacting LLM Networks and Neural SDEs

Vikram Krishnamurthy, Yiming Zhang, Luke SnowarXiv 2026 · arXiv

TL;DR

This paper provides a unified mathematical framework for generative AI by viewing models as stochastic dynamical systems. It bridges the gap between individual token generation in LLMs, the emergent social behavior of interacting AI networks, and the efficient training of continuous-time models (Neural SDEs) using deterministic "cubature" methods instead of slow Monte Carlo simulations.

What Problem Does This Paper Address?

While the AI community often focuses on specific architectures (like Transformers), we lack a rigorous, unified language to describe how these models evolve and interact over time.

The authors address three distinct but related challenges:

  1. The "Black Box" of LLMs: Moving beyond architectural jargon to define LLMs as high-dimensional, nonlinear autoregressive (AR) systems.
  2. AI Social Dynamics: Understanding how "herding," hallucinations, and structural biases (like the "glass-ceiling effect") emerge when multiple LLMs interact in a network.
  3. Continuous-Time Modeling: Improving the efficiency of Neural Stochastic Differential Equations (SDEs), which are essential for modeling irregular data (like financial markets or medical sensors) but are notoriously expensive to train.

Methods

The paper is structured into three "Parts," each introducing a specific mathematical lens:

  • Nonlinear Autoregression: The authors formalize the LLM generation process as a $\Delta$-order AR model. They strip away the complexity to show that the "magic" of LLMs is essentially a repeated composition of bilinear similarity (attention), softmax, and linear aggregation.
  • Mean-Field Dynamics: To study networks of LLMs, they use Mean-Field Ordinary Differential Equations (ODEs). This allows them to predict the "average" behavior of a massive population of agents without simulating every single interaction. They use Random Utility Models (RUM) to approximate how agents choose between being truthful, uncertain, or hallucinating.
  • Wiener-Space Cubature: For Neural SDEs, they introduce a "drop-in" replacement for Monte Carlo estimation. Instead of sampling thousands of random Brownian motion paths, they use a small set of "judiciously chosen" deterministic trajectories (cubature paths) that match the statistical moments of the noise.

Datasets

As this is a tutorial and theoretical paper, it does not introduce a new massive-scale dataset. Instead, it references:

  • Standard LLM Corpora: General discussion of pretraining on 1–20 trillion tokens.
  • Simulated Environments: The authors provide open-source code for simulating the "Collaborative Q&A" and "Glass-Ceiling Effect" experiments.
  • GitHub Repositories: They point to specific implementations for Collaborative Q&A and GCE effects.

Key Insights and Ideas

  • The Illusion of Sentience: The authors argue that the "sentient" appearance of LLMs emerges from the mathematical structure of the bilinear–softmax–linear mechanism composed over many layers, rather than explicit symbolic reasoning.
  • The AI Glass Ceiling: In networks where LLMs autonomously choose who to listen to, "capability-influence mismatches" can occur. Even if an agent is highly capable, structural biases in the network can trap it in a peripheral role, while weaker agents gain disproportionate influence.
  • Deterministic Randomness: The most surprising insight in the SDE section is that you don't actually need "randomness" to train a stochastic model. By using rough-path analysis, you can replace random noise with deterministic paths and achieve better accuracy with fewer computations.

Key Results

  • SDE Training Speedup: The use of Wiener-space cubature improves the root mean-squared error (RMSE) from $O(n^{-1/2})$ (standard Monte Carlo) to $O(n^{-1})$. This means you need significantly fewer "paths" to reach the same level of training precision.
  • Stability Proofs: The authors provide a theorem (Theorem III.1) proving that under certain conditions, a network of interacting LLMs will converge to a unique stable equilibrium regarding its ratio of truthful vs. hallucinated responses.
  • Analytic Gradients: By turning SDEs into an ensemble of ODEs via cubature, the authors show that you can use standard ODE adjoint methods for backpropagation, which are much more memory-efficient than differentiating through stochastic solvers.

Limitations and Future Work

  • Computational Preprocessing: While cubature is faster during training, it requires a preprocessing step to calculate the specific paths and weights for a given "degree" of accuracy.
  • Structural Assumptions: The efficiency gains in Neural SDEs hold only for "suitably structured" drift and volatility functions.
  • Population Limits: The mean-field models assume a "large-population limit." Future work is needed to see how well these ODEs describe very small groups (e.g., a 3-agent debate).
  • Real-World Validation: While the mathematical framework is robust, the authors note that applying these conclusions to "frontier" models (like GPT-4 or Gemini) requires further empirical justification of the "Random Utility" approximations used to model their behavior.

Summary updated

Abstract

This paper develops a stochastic dynamical-systems perspective on three classes of generative AI models. First, we formulate large language models (LLMs) as high-dimensional nonlinear autoregressive models with attention-based dependencies. This formulation connects pretraining by conditional maximum likelihood to sequential token generation and provides a probabilistic interpretation of LLM operation. Second, we consider networks of interacting LLMs and characterize their collective behavior through mean-field ordinary differential equations. These dynamics provide a framework for understanding emergent preferential attachment and persistent structural biases, including type-dependent glass-ceiling effects. Third, we consider neural stochastic differential equations (SDEs) as continuous-time generative models for stochastic trajectories, including those observed at irregular time intervals. We discuss their training with particular emphasis on Wiener-space cubature methods based on rough-path analysis as an efficient alternative to Monte Carlo estimation. Together, the three settings address probabilistic prediction, information exchange and emergent behavior in interacting networks, and path-space generative modeling.
BibTeX
@misc{krishnamurthy2026stochastic,
  author = {Vikram Krishnamurthy and Yiming Zhang and Luke Snow},
  title = {{Stochastic Dynamics of Generative AI: From Autoregressive Models to Interacting LLM Networks and Neural SDEs}},
  year = {2026},
  eprint = {2611.07065},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.07065},
}
Figure