Post-Cutoff.com
  1. Home
  2. Timeline
  3. 2026
  4. Peking University preprint claims an AI-found disproof of…

Peking University preprint claims an AI-found disproof of the Yau–Tian–Donaldson conjecture for constant scalar curvature metrics

★★★★after cutoffsciencePeking UniversityJihao LiuAnthropicOpenAIconfidence: medium

On 19 Aug 2026 Jihao Liu (Peking University) posted a 79-page preprint describing a polarized smooth projective fivefold that is K-polystable but has no constant scalar curvature Kähler (cscK) metric. That would disprove the cscK form of the Yau–Tian–Donaldson conjecture. The paper says 'the main result of this paper was obtained using generative AI', namely Claude Code (Fable 5), Codex (GPT-5.6 Sol) and the Danus research agent. A later solo Danus run re-derived counterexample and proof in 5 h 29 min.

Key facts

Science result

Field
mathematics / Kähler geometry / K-stability
Problem
Yau–Tian–Donaldson conjecture for constant scalar curvature Kähler metrics
Result
Claimed explicit K-polystable polarized smooth projective fivefold without a cscK metric.
AI system
Danus (Rethlas-based agent), Claude Fable 5 (Claude Code), GPT-5.6 Sol (Codex)
Human role
AI-generated main result with one key human redirection; the author checked and polished the text
Verification
Preprint only; not formalised; not peer-reviewed
Status
pending
Why surprising
An agent system is reported to have refuted a flagship conjecture of Kähler geometry, and to have reproduced the result alone in under six hours.

What happened

A Peking University algebraic geometer released a counterexample to the cscK form of the Yau–Tian–Donaldson conjecture. He credits a combination of Claude Code, Codex and his group's Danus agent for the example and its proof. The appendix also records a control experiment: a stronger Danus reproduced the result alone, while six other agent systems could not solve it or even verify the given counterexample.

Why it matters

The Yau–Tian–Donaldson programme is central to modern complex geometry, and the paper is explicit that its main result is AI-generated. The appendix argues that such a counterexample is itself a proof, with no finite certificate, because it asserts something about every possible degeneration. If the result is confirmed, it would be one of the deepest AI-derived results of 2026.

Changelog

  • 2026-09-30: created

Related events

  1. Anthropic releases Claude Fable 5 and Claude Mythos 5 — first generally available Mythos-class model ★★★★★
  2. OpenAI broadly releases GPT-5.6 (Sol, Terra, Luna) after government-gated preview ★★★★
  3. OpenAI's unreleased 'Astra' model claims ten advances in maths and theoretical CS, with Lean proofs ★★★★★
  4. Summer 2026 flood: dozens of named conjectures settled on arXiv with disclosed AI help (July–September catalogue) ★★★★

Sources (2)

id: 2026-08-19-yau-tian-donaldson-cscK-disproof-claim · updated 2026-09-30 · open in the interactive timeline