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

_Published Wednesday, November 11, 2026 at 3:33 AM EST · Science, AI · Latest · Tier 2 — Notable_

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.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2611.13329)

---
Canonical: https://techandbusiness.org/newswire/GQF6ECfFKelydJzLRGMaMK
Published: 2026-11-11T08:33:24.408Z
Story chronology: 2026-11-11T05:00:00.000Z
Retrieved: 2026-11-11T10:24:52.746Z
Publisher: Tech & Business (techandbusiness.org)