Skip to content

Research

GPT-6 Astra helped prove a proposed lattice formulation of the Standard Model

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. Physicist Nathaniel Craig detailed the work on OpenAI’s new science blog.
  2. Astra helped refine the grids and develop the mathematical proof.
  3. The proposal applies at finite lattice spacing. Realistic simulations require further work.
AdvancedDevelopers
A cyan wire lattice with an amber bead against a dark blue background
AI-generated conceptual image: Intersecting metal rods illustrate the idea of placing spacetime on a grid · AI Ping

OpenAI published the first guest post on its science blog on November 15, detailing GPT-6 Astra’s role in a proof for a proposed lattice formulation of the Standard Model of particle physics. Nathaniel Craig, who leads the particle theory group at the University of California, Santa Barbara and is a visiting researcher at OpenAI, said he supplied the strategy and Astra developed the detailed proof.

Craig first submitted the single-author preprint to arXiv on October 6. The paper, its mathematical companion and a formal verification report give physicists materials for examining the assumptions and retracing the argument. The new post explains how the idea moved from a research strategy to a proof and identifies the steps that involved Astra.

The Standard Model describes elementary particles such as quarks and electrons and three forces: electromagnetism, the strong force and the weak force. Lattice methods replace continuous space and time with a grid of points. The grid spacing sets the shortest distance the calculation resolves. Physicists already use this approach to study particles made of quarks and gluons. It gives them a way to include interactions from the outset, without treating each as a small correction, an approach known as nonperturbative. Craig’s project addresses a mathematical obstacle that arises when the framework also has to account for the weak force.

Comparing one grid with the next

The weak force distinguishes between particles of different mathematical handedness. A straightforward lattice rule can introduce extra states of the opposite handedness. A method using overlap fermions can select the desired states, but a phase rule is also needed to combine them consistently. When the force fields change step by step and return to where they started, the states they admit can come back with a different phase.

Martin Lüscher’s earlier work supplied such a construction for an Abelian class of theories under specified conditions. Daisuke Kadoh and Yoshio Kikukawa later worked on a lattice formulation of the electroweak theory. Craig tackles the measure used to combine fermion contributions in the wider non-Abelian class that includes the full Standard Model.

The construction puts the same force fields on a succession of finer grids. Comparing neighboring resolutions yields pieces of a phase correction, whose sum gives a rule on the original grid. The mathematical companion shows why this accounts for the change in phase around each allowed closed path. Craig credited Astra with working out how to represent the fields on the finer grids and developing the proof that the comparisons converge.

Two framed metal grids, the one on the right more closely spaced
AI-generated conceptual image: A coarse grid and a finer one illustrate comparing the same fields at different resolutions · AI Ping

What the proof checks covered

Craig said lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir read the proof and questioned parts of it. The paper’s acknowledgments describe a manuscript combining sections Craig wrote with extensive revisions to Astra’s proof. They also describe tests that Craig and his colleagues proposed, including a check in Lean, software that can verify a mathematical argument expressed in a formal language.

The accompanying report says all 16 modules compiled and that Lean’s kernel separately accepted 95,810 exported declarations, a count that includes library dependencies. It records a separate failed comparison against a reference that required matching compiler-generated proof names. The report documents the results of that comparison alongside the successful compilation and kernel check.

A magnifying glass on a stack of paper with a pencil beside it
AI-generated conceptual image: A magnifying glass over papers represents checking the steps of a proof · AI Ping

The preprint establishes the construction for four-dimensional grids of finite spacing, with a bound on how much the fields can bend across a small square of the grid. A separate two-dimensional supplement applies a related construction and compares phase calculations on specified examples. The examples give readers a concrete check at lower dimension.

Craig is examining whether the construction has a property called reflection positivity at finite lattice spacing. It helps connect the grid description to a physical quantum theory. The mathematical companion also leaves open how to compute the full current efficiently and how to define the interacting theory in continuous spacetime.

Sources