GPT-6 Astra Assists in Formulating Lattice Definition of the Standard Model
Physicists and GPT-6 Astra have formulated a candidate lattice definition of the Standard Model, solving a decades-old theoretical challenge regarding chiral fermions and phase alignment.

For over fifty years, the Standard Model of particle physics has served as the governing framework for describing fundamental particles and three of the universe's four forces: electromagnetism, the weak force, and the strong force. Its theoretical predictions align with experimental data across eighteen orders of magnitude, spanning from the 21-centimeter radio emissions of hydrogen to high-energy particle collisions at the Large Hadron Collider. Despite this empirical track record, the framework has long lacked a complete nonperturbative definition—a mathematical foundation that does not depend on treating interactions merely as small corrections.
In an October preprint, theoretical physicist Nathaniel Craig proposed a candidate definition of the full Standard Model on a spacetime lattice, developed in collaboration with GPT-6 Astra. While Craig formulated the overarching strategy to address the problem, the AI model helped turn that conceptual approach into a concrete mathematical construction and took the lead in producing the detailed proof.
Resolving the chiral lattice problem
Standard calculations in quantum field theory often rely on perturbation theory, assuming particles interact weakly and calculating measurable quantities through successive approximations. While this works accurately for effects like electron magnetism, it fails for phenomena such as the mass of the proton, which arises primarily from nonperturbative interactions among quarks and gluons. To study theories nonperturbatively, physicists discretize spacetime onto a grid or lattice, defining shortest accessible distances and enabling computational simulations.
While lattice field theory successfully explains the strong force, mapping the entire Standard Model onto a grid has remained blocked by chirality. Unlike electromagnetism and the strong force, the weak force treats left-handed and right-handed particles differently. Standard attempts to place chiral fermions on a lattice trigger "doubling," an issue where unwanted mirror partners emerge and spoil the physical particle content.
Prior work by Paul Ginsparg, Ken Wilson, Rajamani Narayanan, Herbert Neuberger, Martin Lüscher, and Yoshio Kikukawa introduced "overlap fermions" to isolate left-handed particles on a lattice. However, setting the correct mathematical phase for the full Standard Model when force field configurations loop back to their start remained an open problem for more than 25 years.
Constructing the proof across finer grids
To overcome this impasse, the construction evaluates force fields from the target grid across a succession of increasingly fine lattices. By measuring the differences between fermion states across adjacent grids and summing those shifts, the method derives the required phase correction for the primary grid.
GPT-6 Astra was instrumental in realizing this refinement framework. The system suggested how to project force configurations onto finer grids, demonstrated that the comparative results would converge to a definite value, and drafted the formal steps of the proof. To verify the result, lattice field theory specialists Maarten Golterman, Andreas Kronfeld, and Yigal Shamir reviewed Astra's proof and provided feedback on its mathematical validity.
The resulting construction provides a candidate nonperturbative formulation at finite lattice spacing under bounded field strengths. Key theoretical questions remain, including whether the formulation satisfies reflection positivity—a mathematical property necessary for conserving quantum probabilities—and how it connects to the continuous spacetime limit, where recent work by OpenAI establishing the mathematical existence of Yang-Mills theory offers a complementary piece.
What it means for developers
This research illustrates a notable shift in how advanced machine learning systems interact with formal sciences. Rather than serving strictly as code generators or text summarizers, frontier foundation models are beginning to handle multi-step mathematical reasoning, helping researchers formalize theoretical strategies and prove convergence criteria.
For developers building automated reasoning tools, symbolic solvers, or scientific workflows, this progress points toward systems that can assist in verifying mathematical constraints and complex software architectures. Developers looking to test and integrate leading frontier models like OpenAI's GPT models, Claude, Gemini, DeepSeek, and Grok into their workflows can access them cheaply via a single API key at https://apixoai.online.
While the current lattice construction is primarily a theoretical advance rather than an immediately deployable algorithm for large-scale numerical simulations, it demonstrates the practical role AI systems can play in resolving longstanding theoretical bottlenecks alongside human domain experts.
Source: A step toward defining the Standard Model on a lattice — OpenAI News. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key

