Atom as an Agent: Emergent Order by Collective Predictive Coding

Koji Hashimoto, Tadahiro TaniguchiarXiv 2026 · arXiv

TL;DR

Researchers have introduced the "Atom-as-an-Agent" (AaaA) principle, applying theories of language emergence to discover physical laws. By treating atoms as "agents" that communicate to predict each other’s states, the team demonstrated that macroscopic order parameters and Landau potentials can emerge automatically from raw data without any prior knowledge of the system's Hamiltonian or symmetries.

What Problem Does This Paper Address?

In physics, "emergence" describes how complex macroscopic behavior (like magnetism) arises from simple microscopic components (atoms). Traditionally, physicists identify the "order parameter"—the variable that describes this transition—by analyzing the symmetries of a system's Hamiltonian (its energy function).

However, this approach requires you to know the underlying physics beforehand. The authors ask: Can we discover these macroscopic rules using only observed data? Furthermore, they seek a deeper information-theoretic reason why these specific variables emerge. They propose that order parameters are essentially a "language" that atoms use to coordinate and predict the collective state of the system.

Methods

The authors utilize Collective Predictive Coding (CPC), a framework originally developed to explain how symbols and languages emerge in human societies.

  1. Atom-as-an-Agent (AaaA): Each atom is treated as an agent with a local probabilistic model. It tries to infer a shared "latent variable" ($w$) that best explains its own state and the states of its neighbors.
  2. Variational Free Energy: The system minimizes an information-theoretic free energy. This process forces the latent variable $w$ to capture the most essential, predictive information about the entire configuration.
  3. Centralized vs. Decentralized Learning:
    • Centralized: A global optimization finds the best $w$ for a given dataset.
  4. Decentralized (Metropolis-Hastings Naming Game): Atoms "talk" to their neighbors, proposing values for the shared variable and accepting or rejecting them based on how well they fit their local observations. This mimics how a consensus "language" might form without a central controller.

Datasets

The study uses synthetic spin configurations generated via Monte Carlo simulations across a range of temperatures for three classic models:

  • 2D Ising Models: Both ferromagnetic (spins align) and antiferromagnetic (spins alternate) versions on a $32 \times 32$ lattice.
  • Frustrated Stripe Model: A 2D spin model where vertical and horizontal patterns compete, testing if the method can identify multiple order parameters ($K=2$).
  • Sherrington–Kirkpatrick (SK) Model: A fully connected spin-glass model used to see if the method can discover "replica overlap," a highly non-trivial order parameter in disordered systems.

Key Insights and Ideas

  • Order as Communication: The "aha moment" is viewing physical order as a communication problem. An order parameter is the most efficient "word" atoms can share to minimize their collective uncertainty about the system's state.
  • Hamiltonian-Blind Discovery: The AaaA principle is entirely data-driven. It doesn't need to know the energy equations or the lattice geometry; it discovers the "uniform" or "staggered" patterns simply because those patterns are the most predictive.
  • Emergent Thermodynamics: Once the order parameter is learned, its probability distribution naturally reconstructs the Landau potential. The researchers showed that the "double-well" shape (indicating two possible states) correctly evolves into a "single-well" shape as the system is heated, matching standard thermodynamic theory.
  • Predictivity as a Metric: The paper argues that a true order parameter must improve the ability to predict hidden parts of a system based on observed parts. They quantified this as "prediction gain."

Key Results

  • High Fidelity: The learned coordinates showed a near-perfect correlation (overlaps $>0.99$) with traditional physical order parameters like magnetization.
  • Automatic Complexity Detection: In the frustrated stripe model, the algorithm correctly identified that two components ($K=2$) were necessary to describe the system, whereas one was sufficient for the standard Ising model.
  • Critical Temperature Accuracy: By analyzing the emergent Landau potentials across different system sizes (finite-size scaling), the authors estimated the critical transition temperature ($T_c$) to be consistent with the known exact values (e.g., $\sim 2.269$ for the 2D Ising model).
  • Decentralized Success: The "Naming Game" successfully allowed local agents to reach a global consensus on the order parameter, proving that macroscopic order can emerge through purely local interactions.

Limitations and Future Work

  • Classical Focus: The current work is limited to classical spin systems. The authors note that extending this to quantum systems—where topological phases and entanglement play a role—is a critical next step.
  • Observation Structure: The method depends on how you define what an agent "observes." For the spin glass, they had to provide "replica pairs" to the agents. Finding the right observation structure for more complex materials remains a challenge.
  • Spacetime Emergence: The authors suggest a bold future direction: applying AaaA to holography and gravity, potentially treating the emergence of spacetime itself as a collective predictive coding problem among microscopic quantum degrees of freedom.

Summary updated

Abstract

What is the intrinsic reason of emergence of macroscopic order parameter out of microscopic systems? In this article, we apply to physics the collective predictive coding (CPC), the theory of language/symbol emergence. As our language emerges out of our human communications, physical macroscopic order parameter emerges out of communications among atoms --- each atom is regarded as an agent who predicts each other's behavior through inference. We present three examples of classical spin systems: two-dimensional ferromagnetic and antiferromagnetic Ising models, a two-dimensional spin model with competing stripe orders, and the fully connected Sherrington--Kirkpatrick spin-glass model. In all of these examples, proper macroscopic order parameters including the replica overlap emerge automatically through communication informatics of CPC, only from observed spin configurations. We call the learned macroscopic coordinate the collective predictive order. Furthermore, for the ferromagnetic and antiferromagnetic Ising models, the equilibrium distribution of the collective predictive order yields an emergent potential with the symmetry and thermal evolution characteristic of a Landau potential. Notably, the emergence of the collective predictive order does not refer to any structure of microscopic Hamiltonians/Lagrangians. Thus our ``Atom-as-an-Agent" (AaaA) principle offers a completely data-driven method for the emergence of macroscopic order, that is, machine learning physics.
BibTeX
@misc{hashimoto2026atom,
  author = {Koji Hashimoto and Tadahiro Taniguchi},
  title = {{Atom as an Agent: Emergent Order by Collective Predictive Coding}},
  year = {2026},
  eprint = {2610.36673},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36673},
}
Figure