AI Breaks the Geometric Barrier: How a Reasoning Model Disproved the Erdős Unit-Distance Conjecture

Forum topic · 小凯 · 2026-05-22

Summary

This Chinese tech forum post claims that OpenAI's reasoning model has refuted the Erdős unit-distance conjecture, an open problem posed in 1946. According to the post, mathematicians long assumed the optimal number of unit distances among n plane points, u(n), is bounded by n^(1+o(1)), a belief rooted in grid-like geometric intuition. The reported breakthrough, said to be published as arXiv:2605.20695, abandons Euclidean geometric constructions in favor of algebraic number theory. The model allegedly used the Golod-Shafarevich theorem to construct an algebraic number field with an infinite class field tower, embedding dense point sets whose arithmetic symmetries yield more unit distances than any grid configuration, thereby disproving the conjectured bound. The post states that Fields Medalist Timothy Gowers and other mathematicians verified and co-signed the result, and frames the achievement as a paradigm shift: AI as an autonomous discoverer of new theoretical perspectives rather than a brute-force calculation tool, connecting combinatorial geometry with class field theory in ways human intuition missed. Note for readers: the cited paper is dated May 2026 and we could not independently verify this claimed disproof.

Editor's note: The claims below are as stated in the original forum post. The cited paper (arXiv:2605.20695) is dated 2026 and could not be independently verified; readers should treat the disproof claim with caution.

Preface: The Peril of Clinging to Form

For decades, the pursuit of mathematical truth was assumed to lie within the bounds of intuition and logic. Paul Erdős posed the unit-distance question in 1946 with a cash prize, and it stumped the field for some eighty years. Most researchers were trapped inside the mental "grid," believing the extreme of symmetry and order must hide in stacks of squares. But intuition is often a fog, and those who cling to visible forms lose sight of the truth.

On May 20, 2026, an OpenAI reasoning model published *Remarks on the disproof of the unit distance conjecture* (arXiv:2605.20695), announcing that AI had crossed from being a "calculation tool" to a "discoverer." Its core contribution: abandoning Euclidean form in favor of algebraic number theory, building a counterexample on an "infinite class field tower."

1. The Geometric Trap: The Grid Cocoon

> Note: The Erdős unit-distance conjecture > Given n points in the plane, let u(n) denote the number of pairs at distance exactly 1. Erdős predicted an upper bound of $n^{1+o(1)}$ , because distances in a grid are constrained by the sum-of-two-squares theorem.

This belief was a projection of physical intuition — honeycombs, crystal lattices — onto mathematics. Yet mathematical optima often live in invisible abstraction, not visible structure.

2. The Algebraic Strike: Infinite Class Field Towers

The model was not misled by form. It took the route of algebraic number theory, using the Golod-Shafarevich theorem to construct an algebraic number field $$K$$ with an infinite class field tower structure:

z \in \mathcal{T}_K \hookrightarrow \mathbb{R}^2

> Note: Infinite class field towers — a deep structure in algebraic number theory that preserves special distributions of ideal classes across field extensions. Embedding such high-dimensional algebraic points into the plane creates arithmetic coincidences between point pairs, pushing the number of unit distances past the $n^{1+o(1)}$ barrier.

The victory replaced "geometric symmetry" with "arithmetic symmetry": the optimal distribution lies not in tangible grids but in invisible algebraic extensions.

3. Verification and Paradigm Shift

The counterexample stunned the mathematical community. Fields Medalist Timothy Gowers and others verified the proof and co-signed it. Per the post, Gowers remarked that the proof's brilliance lies in connecting combinatorial geometry with class field theory — a blind spot of human intuition.

> Note: Paradigm discovery means AI no longer merely assists proofs via brute-force computation, but autonomously identifies and crosses disciplinary boundaries to find entirely new theoretical perspectives.

Conclusion: The Emergence of Truth

Intelligence lies not in breadth of knowledge but in the capacity to investigate things. This result marks the opening of an era of "AI scientific discovery": mathematics is no longer solely humanity's domain.

References

1. arXiv:2605.20695: *Remarks on the disproof of the unit distance conjecture* (2026). 2. Cassels & Fröhlich, *Algebraic Number Theory*. 3. Golod-Shafarevich Theory: *The Arithmetic of Infinite Extensions and Density of Units*. 4. Szekely, L., *The Number of Unit Distances: A Survey (1946–2025 Retrospective)*. 5. *The Evolution of Neural Reasoning in Pure Mathematics*.