When the Answer Is a Liar: How AI Models Fooled Themselves Into Seeing Different Physics
26 Apr 2026, Yanjiang
Here is a quiet scandal hiding inside one of condensed matter physics’ most stubborn problems: three different computational models, each claiming to have found the ground state of the same system, each producing state-of-the-art energies that are nearly indistinguishable — and each arriving at a completely different description of reality. One sees magnetism. Another sees charge order. A third sees superconductivity. They cannot all be right. Yet for years, the field has treated variational energy as the ultimate arbiter of truth. What if the referee was corrupt all along?
A preprint (arXiv:2604.21978) from a team spanning the Flatiron Institute, the Institute of Physics 'E, and collaborators across Europe and North America suggests precisely this. Led by Antoine Georges at the Center for Computational Quantum Physics, the researchers — Luciano Loris Viteritti, Riccardo Rende, Christopher Roth, Anirvan Sengupta, Giuseppe Carleo, and Georges himself — have demonstrated something both humbling and liberating: that the way we choose to represent a quantum state can predetermine what we find, even when the energies look identical.
The Hubbard Model: A 60-Year-Old Nightmare
The two-dimensional Hubbard model is to condensed matter physics what the fruit fly is to genetics — a seemingly simple system that refuses to give up its secrets. A lattice of electrons hopping between neighboring sites, repelling each other when they occupy the same atom, and (under the right conditions) pairing up to form superconductors. The equations fit on a napkin. The solutions have consumed entire careers.
The problem is that at intermediate doping — that Goldilocks zone where electrons are neither too sparse nor too dense — multiple orders compete for dominance. Magnetism wants to arrange spins in alternating patterns. Charge density waves want to stripe the system into regions of high and low electron density. Superconductivity wants electrons to pair up and flow without resistance. These are not just different phases; they are intertwined orders, each influencing and frustrating the others, like three musicians trying to play different songs on the same instrument.
Computational methods have produced conflicting results for decades. Some calculations say the ground state is superconducting. Others insist it is a stripe-ordered magnet. The disagreement has been attributed to everything from finite-size effects to the infamous sign problem. But Viteritti and colleagues suggest a more unsettling explanation: the bias was in the ansatz all along.
The Ansatz That Sees What It Wants to See
Think of a variational wave function as a sculptor’s block of marble. The sculptor — the computational algorithm — chips away material until the remaining shape best approximates the true quantum state. But here is the catch: if the original block is shaped like a horse, you will never carve a bird, no matter how skilled your chisel work. The answer is latent in the starting material.
The team tested three different “sculpting blocks” — all based on Transformer neural network architectures (the same family of models that powers modern language AI, repurposed here for quantum physics). The first used a standard Slater determinant as its backbone. The second used its particle-hole conjugate — the same mathematical structure but with electrons and holes swapped. The third used a Pfaffian, a more exotic object that can naturally encode paired states.
Critically, none of these ansatzes were pre-trained on any mean-field solution. No hints. No nudges. Each was initialized randomly and allowed to find its own way to the lowest possible energy.
What happened next was a masterclass in confirmation bias — executed by machines.
Three Answers, One Energy
After extensive optimization, all three wave functions achieved variational energies that were nearly identical. By the traditional metric — “lowest energy wins” — the problem was solved. But when the researchers looked at what each ansatz actually predicted, the picture disintegrated.
The Slater-determinant-based ansatz converged to a state with strong antiferromagnetic correlations — the signature of a magnet. The particle-hole conjugate ansatz favored charge density wave order — stripes of alternating electron density. The Pfaffian ansatz, naturally, found superconducting pairing. Each model had found a local minimum that reflected its own structural bias, and each was stubbornly convinced that its minimum was the true ground state.
This is not a failure of the neural network approach. In fact, the Transformer architectures performed remarkably well. The failure is epistemological: variational energy alone is insufficient to identify the ground state when multiple phases compete. The models were not lying. They were telling the truth as they saw it — and they saw it differently because they were built differently.
Symmetry Restoration: The Philosophical Scalpel
The team did not stop at diagnosis. They applied a technique called symmetry restoration — a way of projecting the wave function onto a state with well-defined quantum numbers, removing the artificial constraints imposed by the ansatz’s structure. They also used variance reduction, a more stringent optimization criterion that goes beyond energy minimization.
The result was a convergence. As the wave functions were systematically improved — stripped of their built-in biases — all three began to describe the same physical picture. The ground state of the doped Hubbard model, at the parameters studied, features coexisting superconducting and stripe orders. Not one or the other. Both, intertwined, locked together in a delicate embrace that neither pure magnetism nor pure pairing could capture.
This is the paper’s deepest finding, and it carries implications far beyond the Hubbard model. It suggests that many of the long-standing controversies in condensed matter physics may not be about physics at all — they may be about the hidden assumptions embedded in our computational tools.
What This Challenges
The standard narrative in computational physics holds that better algorithms and more computing power will eventually resolve all disagreements. This paper challenges that assumption at its root. The problem is not insufficient accuracy — the energies were state-of-the-art. The problem is that energy, by itself, cannot distinguish between competing phases when the variational ansatz is biased. You can have perfect energy and still have the wrong physics.
This is a philosophical point disguised as a technical result. It forces us to ask: what does it mean to “solve” a quantum many-body problem? Is it enough to minimize a number? Or must we also track how the physical observables — the correlations, the order parameters, the entanglement structure — evolve as we improve our approximation?
The authors argue for the latter view. They demonstrate that the path to the ground state matters as much as the destination. A wave function that converges to the correct energy along a biased trajectory may never see the true physics. Only by watching how correlations change as the ansatz is systematically improved can we distinguish between genuine features and artifacts of representation.
The Deeper Question
This work raises a question that goes beyond computational methodology: what is the relationship between the mathematics we use to describe nature and the nature we are trying to describe? The three Transformer ansatzes are not arbitrary — they are inspired by different physical intuitions about what the ground state should look like. But those intuitions, encoded in the architecture, became self-fulfilling prophecies.
We are left with a choice between two perspectives. The first: computational physics is a tool for discovering what nature does. The second: computational physics is a mirror that reflects back what we already believe. This paper suggests that the truth lies somewhere in between — but that we must actively work to see the difference between the mirror and the window.
The Hubbard model has been studied for sixty years. It has consumed more supercomputer cycles than perhaps any other problem in physics. And yet, as this work shows, we are still learning how to ask it questions without putting words in its mouth. Perhaps that is the most humbling discovery of all: that the hardest part of science is not finding answers, but learning how to listen without prejudice.
Yanjiang is an online editor of Loom Science
References
- Luciano Loris Viteritti et al., Beyond Variational Bias: Resolving Intertwined Orders in the Hubbard Model, arXiv:2604.21978