K3 inside, still irrational?

0:00Cubic fourfolds and K3 surfaces

A cubic fourfold is a smooth hypersurface of degree three in five-dimensional projective space. It is rational if it is birational to projective four-space. Whether a very general cubic fourfold is rational was a famous open question; a 2025 preprint by Katzarkov, Kontsevich, Pantev and Yu claims it is not. But special cubics can be rational. They lie on countably many divisors found by Hassett, labelled by a discriminant d. Pfaffian cubics are rational, and Russo and Staglianò showed that every cubic in the divisors of discriminant 26, 38 and 42 is rational. On these special divisors, K3 surfaces appear. For discriminants that pass a divisibility test, called admissible, Hassett showed that part of the cubic’s middle cohomology is a copy of a polarized K3 surface’s cohomology, shifted in degree and with the sign of the pairing reversed. And Kuznetsov found a piece of the cubic’s derived category, now called the Kuznetsov component. For Pfaffian cubics he showed it is the derived category of a K3 surface, and later work showed the same for every cubic on the admissible divisors. Kuznetsov conjectured that a cubic fourfold is rational exactly when its component is the derived category of a K3 surface. The Hodge-theoretic version predicts the same with associated K3 surfaces.

1:28The claim

Result 054 is a 62-page OpenAI preprint, dated September 24. It claims that for every sufficiently large admissible discriminant, a very general cubic in that divisor is irrational, even though its Kuznetsov component is the derived category of a K3 surface and it has an associated K3 surface in the Hodge-theoretic sense. If it holds, Kuznetsov’s conjecture is false, and an associated K3 surface is not enough for rationality. The threshold is ineffective: the proof gives no way to compute where it starts. Since every cubic of discriminant 42 is rational, it must start above that. The categorical half comes from earlier work of Addington and Thomas, and of Bayer and coauthors. The irrationality is the new part. The paper also derives corollaries: for infinitely many discriminants, these irrational cubics have a Fano variety of lines birational to the Hilbert square of a K3 surface, and all of them have universally trivial zero-cycles, so the decomposition of the diagonal cannot detect their irrationality.

2:38Counting K3 surfaces

The proof counts K3 surfaces along a birational map. Suppose the cubic X were rational, with a birational map f from projective four-space. Fix a collection of K3 surfaces of Picard number one whose transcendental lattice, shifted and with the pairing reversed, matches the cubic’s. For any birational map, count the prime divisors it creates on the target, minus those it removes from the source, each counted only when its maximal rationally connected quotient is a K3 surface from the collection. This count adds up under composition. First, evaluate it with weak factorization: f becomes a chain of blowups and blowdowns along smooth centers. Blowing up a surface adds its whole transcendental lattice to the middle cohomology. Using genus-zero Gromov–Witten pairings, the paper builds an operator algebra that keeps each surface block separate, so a later blowdown must remove a whole block. The cubic has one transcendental block left over, so overall, one more K3 surface from the collection is blown up than blown down. The count is one.

3:47A count that telescopes

Now evaluate the same count with the Sarkisov program: f also factors into links between Mori fiber spaces. The paper constructs integer corrections on these spaces, so that each link’s count is the difference of their values at its two ends. At both ends of the whole chain, projective four-space and the cubic sit over a point, where the corrections vanish. So the total telescopes to zero. The hard part is the conic bundles, where the correction must not depend on how the base is presented. Throughout this half, the link diagrams come in bounded families, which bounds the genus of certain curves. A K3 surface of Picard number one and degree d, the same d as the discriminant, has no pencils of such low genus once d is large. That is where the large discriminant, and the ineffective threshold, come from. One cannot equal zero, so the cubic is irrational.

4:42Where it stands

This is a preprint, not yet peer reviewed, and it has not been revised. There is no Lean formalization, and it cites no other paper from the release. It claims to settle a central conjecture in the field in the negative, so the real test is expert scrutiny of both evaluations: the new Gromov–Witten argument that tracks whole lattices, and the uniform bounds in the Sarkisov half. We found no expert assessment; press coverage so far just repeats the claim. If it holds, a K3 surface in the Hodge structure, or in the derived category, does not guarantee rationality: beyond some unknown discriminant, very general cubics with geometric K3 categories are irrational.