A preprint reports a mixer-and-transformer combination outperforming other neural networks on quantum physics

Simulating a quantum many-body system — a collection of interacting quantum particles — is exponentially expensive, so physicists borrow neural network "ansatze" (learnable shapes used to approximate the quantum state) to make the problem tractable. The standard tool, the Transformer, carries a quadratic compute tax as the system grows: capturing long-range correlations across the lattice gets expensive fast.

A new arXiv preprint from Faith Oyedemi and Stefanie Czischek proposes borrowing an MLP-mixer — a vision-era architecture that mixes information globally and locally without self-attention — as the wave function ansatz for 2D ground states (paper, HTML). On the authors' benchmarks against RNN-based and Transformer-based ansatze, the mixer performs competitively with state-of-the-art models at lower compute, and a hierarchical combination of Transformer plus mixer is reported to outperform other architectures on larger 2D systems.

The result is a real but narrow capability advance: more physics per compute on a benchmark suite the authors defined, not a general scalability proof. The paper is a v1 preprint posted 21 October 2026 with no peer review and no independent replication reported, and "large-scale" refers to the benchmark systems studied, not asymptotic complexity. It is a useful data point in the broader swap of cheaper architectures into neural quantum states, not a shipped system.