Discovering hierarchical governing equations for community-structured complex network dynamics with LLM agents

Jinlei Zhang, Zixuan Zhang, Hao Qiu, Feng Chen, Jiateng Yin, Xiaoyong Yan, Xiaolei Ma, Hao Sun, Jianxi Gao, Lixing Yang, Ziyou GaoarXiv 2026 · arXiv

TL;DR

Researchers have developed SAGE, an LLM-guided agent framework that discovers the mathematical "governing equations" of complex networks by separating shared community behaviors from individual node quirks. Unlike previous methods that rely on fixed libraries of functions, SAGE uses a self-reflective loop to adaptively expand its search space, accurately recovering physical laws even when data is missing or the underlying math is initially unknown.

What Problem Does This Paper Address?

In fields ranging from neuroscience to urban planning, we often observe complex networks—like neurons firing or traffic flowing—and want to know the underlying mathematical rules (governing equations) that drive them. However, two major hurdles have historically blocked the way:

  1. The Hierarchical Mismatch: Most existing tools assume either that every node in a network follows the exact same rule (too simple) or that every node is completely unique (too complex/overfitted). In reality, nodes usually belong to "communities" (e.g., neurons of the same type) that share a core dynamic but have individual variations.
  2. The "Missing Term" Problem: Popular methods like SINDy (Sparse Identification of Nonlinear Dynamics) require a predefined library of candidate functions (like $x^2$ or $\sin(x)$). If the true physical law involves a term not in your library, the discovery process fails.

The authors introduce SAGE (Large Language Model-guided Self-reflective Agent framework for Governing Equation discovery) to bridge these gaps by combining numerical precision with LLM-driven reasoning.

Methods

SAGE formulates equation discovery as a hierarchical problem: it decomposes the dynamics of a node into a community-level law (shared by all nodes in a group) and an individual-level variation (specific to that node).

The framework operates through a "Library–Discovery–Review" loop involving three specialized agents:

  • Library Agent: Manages the candidate functions. Crucially, it is adaptive. If the current model fails, it analyzes the residuals and adds new function families (e.g., trigonometric or exponential terms) to the search space.
  • Discovery Agent: The "math engine." It uses sparse regression solvers (like Lasso or STLSQ) to find the simplest set of terms that fit the data. It identifies shared community structures first, then resolves node-specific residuals.
  • Review Agent: The "critic." It evaluates the discovered equations based on stability, predictive power, and physical interpretability. If an equation is poor, it generates a "repair plan" for the Library Agent to follow in the next iteration.

Importantly, the LLM (GPT-based) does not perform the numerical regression; instead, it acts as a high-level strategist, interpreting why a model failed and suggesting what kind of math might be missing.

Datasets

The authors validated SAGE across a diverse range of systems:

  • Synthetic Networks: A single-dynamics network (Lotka-Volterra predator-prey) and a multi-dynamics network (mixing Lorenz, Rössler, and Chen chaotic systems).
  • Real-World Networks:
    • Urban Mobile Communication: Internet, call, and SMS activity in Milan.
  • Brain EEG: Neural channel oscillations and connectivity.
  • Power Grid: Phase-angle dynamics in the IEEE 39-bus system.
  • Urban Rail Transit: Passenger flow dynamics in the Beijing subway system.

Key Insights and Ideas

  • The "Blind Repair" Capability: In a controlled experiment, the researchers intentionally withheld "sine" terms from the initial library. SAGE was able to look at the "wavy" patterns in the unexplained residuals and correctly reason that it needed to add trigonometric functions to the library to succeed.
  • Disentangling Dynamics: By separating community and individual terms, SAGE provides a "Dynamical Interpretation Report." This makes the results scientifically useful—for example, identifying that subway stations share a "short-term memory" flow pattern but have unique "cross-state interactions" based on their specific urban role.
  • Law Transfer: Because SAGE identifies shared community laws, it can "transfer" knowledge. If one node in a community has zero observations (a sensor failure), SAGE can still predict its behavior using the law discovered from its "neighbors" in the same community.

Key Results

  • Accuracy: SAGE achieved $R^2$ values above 0.99 on synthetic networks, outperforming state-of-the-art baselines like SINDy and PDE-FIND.
  • Robustness: The framework maintained high performance even when 50% of observations were missing.
  • Efficiency: In the "blind repair" test, SAGE recovered the correct equation structure in 90% of runs, while random expansion of the library only succeeded 50% of the time.
  • Interpretability: Unlike black-box neural networks, SAGE produces compact, human-readable equations (e.g., $dx/dt = -12.08y + 5.54$) that scientists can map directly to physical mechanisms.

Limitations and Future Work

While powerful, SAGE has a few areas for growth:

  • Community Dependency: The quality of the discovery depends on how well nodes are clustered into communities. If the initial clustering is wrong, the hierarchical split becomes muddy.
  • Catalog Constraints: Currently, the Library Agent selects new functions from a predefined "catalog" of families. A more open-ended version that can "invent" entirely new mathematical forms is a suggested next step.
  • Stationarity: The model assumes community assignments and governing laws are stable over time. Future versions could tackle "evolving" networks where the rules of the game change mid-stream.

For more details, see the full paper: "Discovering hierarchical governing equations for community-structured complex network dynamics with LLM agents."

Summary updated

Abstract

Governing equations provide an interpretable description of the mechanisms underlying complex network dynamics, yet their discovery from observational data remains challenging when shared community-level dynamics coexist with individual-level variations and the relevant functional structures may be absent from the initial candidate library or not explicitly specified in advance. Here we formulate this task as a hierarchical law-discovery problem and introduce SAGE, a large language model-guided self-reflective agent framework for discovering community-shared governing laws and individual-specific dynamical variations. SAGE organizes equation discovery into an iterative Library-Discovery-Review process that first identifies shared community-level structures and then resolves node-specific residual dynamics, while adaptively refining the candidate space to introduce missing dynamical terms when needed. This hierarchical and interpretable formulation enables SAGE to distinguish shared community-level dynamics from individual-level variations and to reveal the mechanisms underlying complex network dynamics. Across two synthetic networks and four real-world networks spanning mobile communication, brain electrophysiology, power grids and urban rail transit, SAGE recovers compact and interpretable governing equations, reproduces observed system dynamics and outperforms representative equation-discovery baselines. The framework remains robust under incomplete observations and can transfer community-level governing laws to nodes with entirely missing observations. These results establish a hierarchical and adaptive approach for disentangling shared and individual dynamical mechanisms in community-structured complex networks.
BibTeX
@misc{zhang2026discovering,
  author = {Jinlei Zhang and Zixuan Zhang and Hao Qiu and Feng Chen and Jiateng Yin and Xiaoyong Yan and Xiaolei Ma and Hao Sun and Jianxi Gao and Lixing Yang and Ziyou Gao},
  title = {{Discovering hierarchical governing equations for community-structured complex network dynamics with LLM agents}},
  year = {2026},
  eprint = {2611.02816},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02816},
}
Figure