Talagrand's critical Sherrington-Kirkpatrick overlap conjecture
At the critical inverse temperature in the Sherrington-Kirkpatrick spin glass model, Talagrand conjectured that the expected squared overlap of two independent Gibbs replicas has an exact scaling: there exists a constant such that
Du and Huang prove that this limit exists and is positive and finite. More strongly, they determine the full limiting quenched distribution of the rescaled overlap in terms of the reflected point process.
- Result
- Proved(see note)
- Status
- Resolved
- AI contribution
- AI co-developed
- Method
- Argument
- Field
- Spin glass theory
- Posed by
- Michel Talagrand
- Year posed
- 2011
- Years open
- 15y
- Solved
- 2026-08-09
- Model
- GPT-5.6 Pro
- Vendor
- OpenAI
- Collaborators
- Hang Du, Brice Huang
- Verification
- Unreviewed
- Publication
- Preprint
- Significance
- 30 / 100
- Disclosed cost
- —
- Wikipedia
- No dedicated article
What was actually shown
Talagrand's Conjecture 11.7.5 is resolved affirmatively for the critical Ising Sherrington-Kirkpatrick model: converges to a positive finite constant. The paper proves substantially more, showing that the entire quenched distribution of converges to an explicit random probability measure defined from the reflected point process. The same limiting distribution and second-moment constant are obtained for the spherical SK model. This does not resolve the broader low-temperature overlap structure of the SK model.
What the AI did
The authors state that most of the arguments in the paper were generated using GPT-5.6 Pro, with the goal of exploring further consequences of ideas developed in their companion work on critical SK free-energy fluctuations. The paper does not assign individual lemmas or the central comparison principle specifically to the model, so the contribution is classified conservatively as AI co-developed rather than AI discovered.
Verification
Checked by this site on 14 August 2026 against the paper's LaTeX source and the prior-art source, both fetched and read. Confirmed verbatim in the paper: "The following corollary affirmatively resolves [Conjecture 11.7.5]", attached to the corollary giving as the mean of an explicit positive random variable built from the reflected process. The prior state is independently pinned: Dey and Kang (arXiv:2603.05636, March 2026) quote Conjecture 11.7.5 with exactly the statement submitted here and treat it as open. The AI disclosure was read in full: most formal arguments were initially generated by GPT-5.6 Pro, the principal human inputs are named (identification of the limiting objects, formulation of the sphere-to-cube comparison theorem, and the overall proof strategy), and the technical body separately credits one step to the model by name - the Hilbert-space kernel argument upgrading the Laplace-transform comparison is introduced as "an idea due to GPT-5.6 Pro". What was NOT checked here is the mathematics itself: a 36-page argument through the scaling limit at the GOE spectral edge needs a spin-glass or random-matrix specialist. The manuscript is five days old, unrefereed, and the authors taking responsibility for correctness is author-side checking, so the tier stays Unreviewed.
Sources
Submitted by HiddenHawk615 on