Continuum neural wavefunctions for anyons on a torus

Mobile anyons in topologically ordered phases provide a setting to investigate how fractional exchange statistics shapes collective behavior at finite density. We introduce AnyFormer: a neural variational framework for continuum anyons on a torus that explicitly incorporates statistical interactions and global winding constraints. We construct a smooth orbital frame that enforces the matrix-valued torus boundary conditions, and supplement it with an analytic pair factor that supplies the regular collision behavior. The orbital geometry and many-body amplitudes are optimized in parallel by variational Monte Carlo. We apply AnyFormer to semions without an external field or interaction potential, isolating unique correlations generated solely by statistics. Our framework closely matches exact two-semion energies and yields lower variational energies than finite-cutoff exact diagonalization for $N = 4$ and $6$ semions. Scaling up to $N = 16$ semions, well beyond the limits of continuum diagonalization, our calculations reveal states with pronounced correlation holes, suppressed long-wavelength density fluctuations, and a reversal of pair circulation at a distance set by the mean particle spacing. For $N = 12$ and $16$ semions, the circulation develops damped spatial oscillations. AnyFormer thus provides a scalable variational neural-net framework for microscopic studies of correlations, chiral motion, and collective phases of bulk anyonic matter.