AI agents find physics functional with 11.7% lower held-out test error

Researchers report in a preprint that AI agents developed a 13-parameter formula for electronic-structure calculations that lowered held-out MGCDB84 test error by 11.7% against ωB97M-V. The formula, O159, retained its improvement when evaluated self-consistently, a test of whether its gains survive use within the calculation itself. The agents worked through a GPU evaluator to propose and fit formulas. O159 adjusts same-spin electron correlation through a smooth density window. Only one candidate advanced to self-consistent evaluation; across 36 GSCDB137 datasets, its mean normalized error ratio was 1.139, comparable to ωB97M-V's 1.144.