Physicists Train AI to Discover Its Own Order Parameters From Spin Data
A new paper treats atoms as agents chatting through Bayesian inference, letting a neural model discover order parameters from raw spin snapshots alone.
Takeaways
- New preprint applies Collective Predictive Coding, a language-emergence theory, to discover physical order parameters from raw data.
- Each atom is modeled as an agent predicting its neighbors through decentralized Bayesian inference.
- Method recovers magnetization, staggered magnetization, stripe sectors, and SK replica overlap without ever seeing the Hamiltonian.
- Equilibrium distribution of the learned coordinate reconstructs a Landau-style double-well potential across the Ising transition.
- Finite-size extrapolation of the inferred critical temperature converges toward the exact Onsager value.
- Framework ports Taniguchi's CPC hypothesis from cognitive science into statistical physics; code not yet released.
Physicists use order parameters to connect microscopic configurations with macroscopic phases. Magnetization identifies a ferromagnet, staggered magnetization identifies an antiferromagnet, and replica overlap measures shared structure in a spin glass. Researchers usually select these quantities after inspecting the Hamiltonian, the function that defines a system’s interactions and energy. In the preprint Atom as an Agent, Koji Hashimoto and Tadahiro Taniguchi ask whether a learning system can infer the appropriate collective coordinate from spin configurations alone. Their method treats each spin site as an agent whose shared representation must help predict other sites.
A language game for spins
Collective Predictive Coding, or CPC, originated as a cognitive-science account of symbol formation. Taniguchi’s 2024 CPC paper describes agents that maintain separate internal models while exchanging proposed symbols. A Metropolis-Hastings Naming Game supplies the computational mechanism: one agent proposes a symbol, another accepts or rejects it according to how well it explains local observations, and repeated exchanges approximate Bayesian inference over a shared representation.
Applied to spin data, each site becomes an agent and the shared latent variable w becomes a candidate order parameter. Its value must carry information that helps one site predict another. Hashimoto and Taniguchi call the resulting variable the collective predictive order.
Three systems, known answers
The authors test CPC on classical spin systems with established order parameters, allowing them to compare the inferred coordinate against a known target.