Structural benchmarks overlook density-of-states errors in machine-learned interatomic potentials

Universal machine-learned interatomic potentials are widely used as fast surrogates for density functional theory in structural modeling. Standard benchmarks emphasize energies, forces, and relaxed geometries. Whether the resulting structures also preserve the density of states remains largely unexplored. We examine this link across oxides, iodides, and alloys by relaxing structures with four machine-learned interatomic potentials, recalculating the density of states with a common density-functional protocol, and comparing against fully density-functional references. Across all three material classes, structural benchmark performance does not predict electronic agreement. Nevertheless, the hybrid workflow remains useful: single-point density-functional calculations on machine-learned geometries recover Fermi energies, reduce convergence cost, and approximate the ensemble-averaged density of states of a disordered high-entropy alloy from a small subset of ordered representatives. These results show that structural benchmarks alone are insufficient for assessing density-of-states agreement and identify a practical route for reducing the cost of density-functional workflows.