The Standard Model of particle physics agrees with observations to high precision across 18 orders of magnitude in energy, from the 21 cm radio waves emitted by hydrogen atoms to high-energy experiments at the Large Hadron Collider (LHC). Yet more than 50 years after its birth, a fundamental challenge remains: a "non-perturbative definition" of the whole theory, one that does not treat interactions between particles as small corrections.

On October 6, 2026, Nathaniel Craig, a theoretical physicist at the University of California, Santa Barbara, released a preprint prepared with the help of the AI model "GPT-6 Astra": "A gauge-invariant measure for lattice chiral fermions" (arXiv:2610.09093). On November 15, Craig himself explained the work on OpenAI's website. The paper proposes a mathematical construction needed to define the Standard Model non-perturbatively on a lattice. Below, we look at what solution it offers to a long-standing open problem and what remains before the theory is complete.

AD

The Limits of Perturbation Theory, and Why Define the Standard Model on a Lattice

When physicists derive concrete numbers from the Standard Model, they often use a method called "perturbation theory." One first solves a simple system with no interactions, then adds the effects of interactions step by step as small correction terms. When the coupling is weak, as in electromagnetism, computing just a few correction terms yields extremely high precision. For example, the magnetic moment of the electron is predicted by perturbation theory to a precision of a few parts in a trillion.

In regimes where interactions are strong, however, such approximations do not work well. A prime example is the mass of the proton. A proton is made of three quarks, but most of its mass comes not from the quarks' own masses but from the energy of the motion and interactions of the quarks and the gluons that mediate the strong force. Without the strong interaction the proton as a bound state would not exist, so a method that treats interactions as small corrections cannot yield its mass. Handling such phenomena requires non-perturbative methods that include interactions from the start.

For the Standard Model as a whole, no established way exists to define the theory non-perturbatively in continuous spacetime, and whether an interacting theory is mathematically consistent is not fully settled. One leading approach is "lattice field theory," which replaces continuous spacetime with a grid of points at regular spacing. By treating the lattice spacing as the shortest distance in the theory and defining the interacting theory on a finite lattice, one can carry out analyses and computer calculations that do not rely on perturbation theory.

Indeed, lattice quantum chromodynamics (lattice QCD), which handles the strong interaction, has delivered results over the past several decades, including a quantitative first-principles determination of the proton mass. Extending the approach to the full Standard Model, including the weak interaction, has run into a different kind of mathematical barrier. What Craig tackled is the definition of the "integration measure" required when handling left-handed fermions on a lattice.

From Fermion Doubling to a Phase Problem Open for Over 25 Years

The difficulty in extending lattice field theory to the whole Standard Model is that the Standard Model is a "chiral theory" that distinguishes left from right. Just as a left-handed glove becomes a right-handed one in a mirror, fermions such as quarks and electrons come in left-handed and right-handed types. The strong and electromagnetic interactions treat the two the same way, but the weak interaction treats left-handed and right-handed fermions differently. This asymmetry is a key feature of the Standard Model.

However, placing chiral fermions on a lattice naively produces unintended extra particles, a problem known as "fermion doubling." Even when one tries to represent only left-handed fermions, extra right-handed degrees of freedom appear, and the particle content inherent to the weak interaction cannot be reproduced.

Researchers have developed solutions to this problem step by step. In 1982, Paul Ginsparg and Kenneth Wilson derived a condition for handling chiral symmetry properly on a lattice (Physical Review D paper). Known as the "Ginsparg-Wilson relation," this condition became the theoretical foundation for treating chiral symmetry while avoiding doubling.

Rajamani Narayanan and Herbert Neuberger then developed the "overlap fermion" construction, in papers including one published in 1994. This gave a concrete way to handle left-handed fermions on a lattice while removing unwanted particle degrees of freedom.

Method / Year Theory / Gauge group Main result Remaining issue
Ginsparg-Wilson (1982) Chiral symmetry on the lattice Presented a modified condition for handling chiral symmetry on a lattice A concrete fermion construction and a gauge-invariant measure are needed separately
Narayanan-Neuberger (1994) Lattice chiral gauge theory Developed the overlap fermion construction The phase of the gauge-field-dependent integration measure unresolved
Lüscher (1999) Anomaly-free abelian gauge theory Constructed a smooth, local, gauge-invariant measure Not applicable to general non-abelian gauge theories
Yoshio Kikukawa et al. (2001) Electroweak theory Proved global integrability for gauge loops of Proof for non-gauge loops limited to classical $SU(2)$ instanton backgrounds
Craig (2026, preprint) The four global gauge groups of the Standard Model Presents a construction of a smooth, local, gauge-invariant measure on a finite lattice Limited to admissible gauge fields; reflection positivity and the continuum limit unresolved

Even when overlap fermions remove the extra particles, the question of how to fix the "phase" of the integration measure remains. In quantum field theory, physical quantities are computed by integrating over all field configurations. The measure used for this integration must have the property that gauge transformations do not change physical results and that it does not depend unnaturally on fields far away.

However, when one extracts the left-handed fermion degrees of freedom, an ambiguity arises in the phase of the integration variables, depending on the background gauge field. If one changes the gauge-field configuration continuously and returns it to the original configuration after a full circuit, the phase of the integration measure may fail to return to its original value, even though the allowed fermion states are the same. Resolving this requires defining the phase so that it is consistent over all allowed closed paths.

In 1999, Martin Lüscher showed how to fix this phase consistently for anomaly-free abelian gauge theories with gauge group , such as a chiral version of electromagnetism (original paper). In 2001, Yoshio Kikukawa and colleagues proved global integrability for loops corresponding to gauge transformations, for the gauge field of electroweak theory. For loops not arising from gauge transformations, however, the proof was limited to classical $SU(2)$ instanton backgrounds (preprint).

Partial constructions for non-abelian gauge theories have been studied since, but problems remained unresolved in a non-perturbative measure construction covering the whole Standard Model. Craig's work does not invent a new way of constructing chiral fermions. Its distinguishing feature is that it constructs a gauge-invariant integration measure for the existing overlap fermions under specific mathematical conditions.

AD

A New Proof That Refines the Lattice and Cancels Phase Mismatches

The core of Craig's method is "refinement": mapping gauge-field configurations onto successively finer lattices and accumulating the differences.

First, the gauge field on the original lattice, on which one wants to define the theory, is represented on lattices with the spacing successively halved, then halved again. Then the projection operators that select the allowed chiral fermion states are compared between adjacent coarse and fine lattices. This comparison yields a contribution at each stage to the "measure current," which describes the change in the integration measure.

When the gauge field is varied along a closed path, a phase change of the fermions arises on each lattice. But when the differences between adjacent lattices are summed, the phases corresponding to the intermediate lattices cancel one after another. This kind of calculation is called a "telescoping sum." If one can show that the residual phase on the infinitely fine lattice converges to 1 and that the series for the measure current converges, the phase change arising on the original lattice can be corrected consistently.

Key to proving this convergence is a property of the Standard Model called "anomaly cancellation." An anomaly is a phenomenon in which a symmetry that holds in the classical theory is broken by quantum effects. If a gauge anomaly remains, the consistency of the theory is compromised. In the Standard Model, the contributions of the fermions combine so that the required anomalies cancel, and this property underpins the convergence of the series obtained through refinement. In addition, to control the global phase, the paper introduces an auxiliary $SU(3)$ connection accompanied by conjugate fermion pairs.

Using these constructions, Theorem 4.1 of the paper asserts the existence of a measure current that is smooth, gauge covariant, and whose dependence on distant gauge fields decays exponentially. This current satisfies the covariant divergence equation for the gauge anomaly and the complete integrability condition for all admissible closed loops. As a result, the paper states, a smooth, gauge-invariant fermion integration measure can be constructed on each nonempty connected component of the space of admissible gauge fields.

However, the theorem concerns a finite, periodic four-dimensional Euclidean lattice, and the gauge fields are subject to a condition called "admissibility." Specifically, the product of the gauge field around a plaquette, the smallest closed loop on the lattice, must be sufficiently close to the identity matrix. This condition allows the properties of the overlap operator to be controlled appropriately.

On the other hand, the threshold that defines admissibility and the decay rate of locality do not depend on the lattice volume. Nor is it necessary to assume that individual link variables are close to the identity or that interactions are weak. The proof is a mathematical construction on a finite lattice; it is not a study that computed particular physical quantities through numerical simulation.

Covering the Standard Model's Four Gauge Group Structures, but the Measure Is Not Fully Unique

Craig first took up the $SU(5)$ gauge group, also used in grand unified theories, and constructed the integration measure for left-handed fermions belonging to its and \bar} representations. This combination corresponds to one generation of Standard Model fermions.

He then showed how to map link variables belonging to the Standard Model gauge group into the $SU(5)$ construction and apply the resulting measure to the Standard Model. The paper treats all four possible global structures of the Standard Model gauge group. These correspond to and its quotients by discrete subgroups. The same construction can also be extended to multiple fermion generations, so it covers the Standard Model's three generations as well, according to the paper.

However, the constructed integration measure is not completely unique. If the space of admissible gauge fields splits into several disconnected components, the freedom remains to choose a constant phase for each component. Gauge invariance alone cannot fix the relative phases between different connected components, and depending on the choice of phases, some discrete symmetries may not be preserved.

The paper also shows cases where the construction extends to more general chiral gauge theories, but additional conditions apply. Specifically, in addition to the usual local anomaly cancellation, half-integer isospin representations of a certain $SU(2)$ subgroup must appear with even multiplicity. This condition is used to make the argument about global phases work, so the proof does not apply to every anomaly-free chiral gauge theory.

What the paper shows is that, under these conditions, a fermion integration measure exists and has the required mathematical properties. It does not calculate new particle masses, scattering cross sections, or predicted values for experimental results.

AD

How Did GPT-6 Astra Take Part in the Proof?

The work also draws attention as a case of AI taking part in the mathematical proof of advanced theoretical physics. However, one must distinguish between the researcher's own account of which parts the AI handled and independent academic evaluation.

According to Craig's explanation on OpenAI's site, the basic strategy of successively refining the lattice was originally Craig's idea. But he had been unable to prove that the method works in all the necessary cases. He therefore used GPT-6 Astra to move forward on making the construction concrete.

In Craig's account, Astra proposed an appropriate way to map the gauge field on the original lattice onto finer lattices and showed that the series obtained from comparisons between lattices converges to a definite answer. It also played a central role in assembling the detailed proof. Craig states that Astra's support was indispensable in developing the refinement idea into a mathematical construction that actually works.

Human experts were also involved in examining the proof. According to the OpenAI article, Maarten Golterman, Andreas Kronfeld, and Yigal Shamir, who specialize in lattice field theory, read the proof and raised technical questions and points. Craig explains that these discussions were important in deepening his confidence in the result. However, informal discussions with experts differ from formal peer review at a journal, and the published preprint is, at this point, research that has not yet been peer reviewed.

The distinguishing feature of this result is that the author reports that the AI was involved not merely in searching existing knowledge or polishing text, but in proposing a mathematical construction and elaborating the details of the proof. At the same time, this division of labor rests on Craig's own explanation; it is not an independent experimental evaluation of the AI's mathematical reasoning ability.

Reflection Positivity and the Continuum Limit Remain Before the Standard Model Is Complete

Even if this mathematical construction is correct, it does not by itself complete a non-perturbative definition of the Standard Model. One important remaining issue is "reflection positivity" at finite lattice spacing.

Reflection positivity is a condition for connecting a quantum field theory defined in Euclidean spacetime to a quantum theory with ordinary time evolution. Satisfying it is important for building a theory in which the probabilities of physical states are properly preserved. For Craig's measure construction, it has not yet been shown that this property holds.

Another issue is the "continuum limit," in which the lattice spacing is taken toward zero. Even if a theory can be defined mathematically on a finite lattice, that does not guarantee that, as the lattice spacing is reduced, an interacting quantum field theory in continuous spacetime emerges properly. Craig's paper also leaves the definition of the interacting continuum theory as an open problem.

Furthermore, to use this method in actual lattice simulations, it will need to be developed into a computable form. Craig himself explains that the construction is closer to an answer in principle than a practical computational method. Even if an integration measure can be constructed in theory, it does not follow that one can use it to compute Standard Model quantities efficiently.

In the same article, OpenAI also mentions work on an existence proof for Yang-Mills theory, which is being pursued as a separate research effort. Such work touches on the shared challenge of establishing non-perturbative quantum field theory, but Craig's paper itself does not prove the existence of the Standard Model in continuous spacetime.

In assessing the result, the first thing that matters is whether Theorem 4.1 and its supporting proofs withstand detailed scrutiny by experts. Beyond that, the next stages of research will be clarifying the conditions under which reflection positivity holds, connecting to the continuum limit, and applying the construction to numerical computation. For the long-standing problem of the integration measure on a finite lattice, the construction presented by Craig and GPT-6 Astra offers new material for the non-perturbative understanding of the Standard Model.