SpinGlassBench: Exact References Expose Hidden Errors in Neural Samplers and Free-Energy Estimators
Year: 2026
Type: preprint
Abstract: Exact ground truth for neural samplers and free-energy estimators. Learned samplers (autoregressive networks, discrete diffusion, GFlowNets) are often evaluated on systems where nobody knows the right answer. SpinGlassBench uses the Sherrington-Kirkpatrick (SK) spin glass, where the exact ground truth is computable by exhaustive Gray-code enumeration up to n = 28 spins (268,435,456 states per instance) and in the infinite-volume limit via the Parisi formula and the Aizenman-Lebowitz-Ruelle (ALR) high-temperature theorem. This release package contains:- SpinGlassBench_Manuscript.pdf: Full 20-page manuscript draft.- SpinGlassBench_Supplementary_Material.pdf: Detailed mathematical derivations, ablations, and extensive tables.- spinglassbench_code.zip: Full reproducible codebase (Rust Gray-code enumeration engine, PyTorch training harness, evaluation scripts, and test suite).- spinglassbench_benchmark_data.zip: 1,400 disordered instances, exact partition functions, reference ledger, and full benchmark records across classical (PT-TI, SA-TI, AIS, SMC, PA) and learned samplers (VAN-MADE, discrete diffusion, GFlowNet, transformer).- Checkpoints & Audits: Traced verification of released checkpoints from MDNS, MetaDNS, and LEAPS exposing numerical precision and discretization errors.
Author: Gangadharaiah, Karthik
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225335971