VibeMathedMath problems solved by AI

Seymour's Second Neighborhood Conjecture

Seymour conjectured that every finite oriented graph has a vertex with at least as many exact second outneighbors as outneighbors. Known cases include tournaments (Fisher 1996) and minimum outdegree at most six (Kaneko-Locke 2001), and for dense incomplete graphs a series of results restricting the structure of the missing edges. This work proves the conjecture for every oriented graph of order n=2δ+2n = 2\delta + 2, where δ\delta is the minimum outdegree, with no prescribed structure on the missing edges; with Fisher's tournament theorem this gives every oriented graph satisfying n2δ+2n \le 2\delta + 2.

Result
Proved(see note)
Status
Partial result
AI contribution
AI-discovered
Method
Argument
Field
Graph theory
Posed by
Paul Seymour
Year posed
1990
Years open
36y
Solved
2026-08-12
Model
GPT-5 family, Claude
Vendor
OpenAI, Anthropic
Collaborators
Jake Brukhman
Verification
Unreviewed
Publication
Preprint
Significance
30 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

A dense case, not the conjecture: it remains open in general. The concrete gain is on the size of any counterexample - combined with the known minimum-outdegree results, this raises the best known lower bound on the order of a counterexample from 16 to 17, and to 19 conditional on the 2026 preprint of Sadhukhan, Sandeep and Sen. The novelty against the earlier dense-case work is that no structure is prescribed on the missing edges.

What the AI did

The acknowledgements are unusually specific about the division of labour: "The author initiated and directed the investigation, curated intermediate results, selected the theorem for publication, and edited the final statement and exposition. OpenAI language models (GPT-5 family) carried out the detailed mathematical exploration, implemented counterexample searches and verification tools, discovered the fixed-target capacity argument and its double-counting proof, and drafted the manuscript; Anthropic Claude models performed an adversarial audit of an intermediate draft and assisted with revisions. The author verified the proofs and accepts sole responsibility for the final manuscript and its claims." The model is credited with discovering the central argument by name, which is the discovered tier rather than the co-developed one.

Verification

arXiv preprint, one day old at cataloguing, not peer-reviewed and with no independent commentary yet. The proof is described by the author as a short counting argument, so it is human-checkable in principle, but this site has not verified it and no expert has publicly endorsed it. What is unusually strong here is the disclosure rather than the verification.

Sources

Changelog1 change
  • Rasmus Lindahladded this entry

Discussion