Erdős–Sós Conjecture: AI-Assisted Proof Shakes Math
Mathematicians are buzzing over a new arXiv preprint claiming an AI-assisted proof of the Erdős–Sós conjecture, a landmark open problem in graph theory dating back over 50 years. If confirmed, it would mark one of the first major classical conjectures cracked with substantial help from machine reasoning. Researchers posted the paper (arXiv:2609.17877) this month, and the community is already calling it a potential turning point.
What is the Erdős–Sós conjecture?
The conjecture, posed by Paul Erdős and Miklós Sós in the 1970s, asks a deceptively simple question: how large can a graph be before it is guaranteed to contain every tree of a given size as a subgraph? It has survived decades of partial results and resisted attack by some of the field's best minds. A full proof has long been seen as one of extremal graph theory's holy grails.
How much of the proof is actually AI?
The preprint describes a proof where AI systems generated key lemmas and structural insights that human mathematicians then verified and assembled into a formal argument. The authors also report that parts of the reasoning were checked in the Lean proof assistant — the same tool behind other recent AI math pushes. This looks like a collaboration model, not a fully autonomous proof, but even that would be a first at this scale.
What does this mean for AI research?
The paper arrives alongside a separate arXiv study testing AI systems on 68 of the hardest open Erdős problems in a Lean-verified setting, suggesting a broader shift: AI as a genuine research partner in pure mathematics, not just a solver of olympiad-style exercises.
Caveat: this is a preprint, so it has not yet passed peer review. Experts will likely scrutinize the proof line-by-line, and a subtle gap could still sink the result. Until then, treat the claim as exciting but provisional — the community reads it as a breakthrough, but verification is what actually crowns it.