VibeMathedMath problems solved with AI

The Hodge conjecture for products of K3 surfaces

The rational Hodge conjecture asserts that on a smooth projective complex variety every rational (p,p)(p,p) class is algebraic. For a single K3 surface this is the Lefschetz (1,1)(1,1) theorem, but products S1×⋯×SmS_1\times\cdots\times S_m carry Hodge classes coupling the transcendental cohomologies of the factors, and powers carry further invariant tensors. Known cases include Hodge isometries (Mukai, Nikulin, Buskin), CM endomorphism fields (Ramon Mari), and special families via algebraic Kuga-Satake correspondences (Schlickewei, Varesco, Floccari). Does the rational Hodge conjecture hold in every codimension for every finite product of projective complex K3 surfaces?

Result
Proved(see note)
Status
Retracted
AI contribution
AI-discovered
Method
Argument
Field
Algebraic geometry: Hodge conjecture, K3 surfaces
Posed by
W. V. D. Hodge (1950) for the conjecture; the K3 product case is a long-studied special case (Mukai, Ramon Mari, Varesco, among those cited)
Year posed
1950
Years open
76y
Solved
2026-09-10
Model
Unreleased internal OpenAI model
Vendor
OpenAI
Collaborators
—
Verification
Contested
Publication
Announced
Collection
OpenAI math release (October 2026), version fd4aeeb
Significance
50 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Claims the rational Hodge conjecture in every degree for every finite product of projective complex K3 surfaces, with no restriction on Picard numbers, periods or endomorphism fields, including powers of one surface. On the way it re-proves algebraicity of the exact Kuga-Satake correspondence. It says nothing about the integral Hodge conjecture or about hyperkahler varieties beyond what follows.

What the AI did

The OpenAI math release (github.com/openai/math) states its results were produced by an unreleased internal OpenAI model, with on average about three hours of ChatGPT Pro thinking compute per result, under one fixed procedure applied to roughly 4,000 posed problems; outputs were then grouped into families and filtered for significance. The manuscript is credited to OpenAI alone and names no human author. The proof uses the companion CM Hodge theorem as an input, and the README lists that theorem as produced outside the fixed procedure.

Verification

No independent mathematician has checked this yet. The release withdrew this manuscript on 6 October 2026, with the Kuga-Satake paper and the split abelian eightfold paper it builds on: the proof is not established (see the claim issue). Checked here: Theorem 1.1 of the TeX source read against the posed problem; surjectivity of CHp(X)Q→Hdgp(X)CH^p(X)_\mathbb Q\to \mathrm{Hdg}^p(X) for every product of projective K3 surfaces, repeated or distinct, every pp. Inputs named by the paper and not checked here: the companion quadratic-locus criterion, the companion CM Hodge theorem, Buchweitz-Flenner and Pridham semiregularity, and homological mirror symmetry for the quartic and tori. The mixed-product step is new. No Lean formalization is listed.

Claim issue

Withdrawn by the authors. OpenAI's withdrawal notice of 6 October 2026: "The proof relies on an adaptation of the flawed stabilization-trace construction in Algebraicity of Weil classes on split abelian eightfolds. The sign error in that construction prevents the required signed-double-point cancellation. The resulting gap leaves this paper's claimed rational Hodge theorem for products of projective K3 surfaces unproved." The notice adds that the withdrawal concerns the proof and does not assert that the statement is false. The question is open again.

Sources

Changelog4 changes
  • Rasmus Lindahlchanged Verification from “unreviewed” to “contested”, also Verification note, Claim issue, collectionVersion
  • Matarisvanset Significance note to “A natural and much-studied case of the Hodge conjecture beyond abelian varieties, known be…”, also What the AI did, Status, More links, Verification note
  • QuietNarwhal605commented
  • Rasmus Lindahladded this entry

Discussion2

QuietNarwhal60508 Oct 2026, 03:29 UTC· edited
1
Rasmus Lindahl08 Oct 2026, 14:46 UTC

Thanks for flagging it. OpenAI withdrew this manuscript on 6 October, together with the Kuga-Satake paper and the split abelian eightfold paper it builds on: a sign error in the eightfold stabilization-trace argument breaks the signed double-point cancellation the construction needs. The entry is now Retracted and Contested, with OpenAI's notice quoted in the claim issue and linked. As their notice says, the withdrawal concerns the proof; it does not show the statement is false, so the question is open again.

0