AI Breaks New Ground in Ramsey Theory with Improved Lower Bounds

A team of researchers has achieved significant progress in Ramsey theory, a branch of combinatorics that studies conditions under which order must appear. Using AlphaEvolve, an LLM-based code mutation agent, they've improved lower bounds for five classical Ramsey numbers, marking a notable advance in a field that has seen incremental progress for decades.

What Are Ramsey Numbers?

Ramsey numbers are fundamental objects in combinatorics that answer questions about unavoidable patterns in large structures. The Ramsey number R(s,t) represents the smallest number of vertices needed in a complete graph such that any coloring of its edges with two colors will necessarily contain either a clique of size s in the first color or a clique of size t in the second color.

For example, R(3,3) = 6 means that in any group of six people, either three mutually know each other or three are mutual strangers. Despite their simple definition, Ramsey numbers grow extremely rapidly and are notoriously difficult to compute.

The New Results

The researchers improved lower bounds for five specific Ramsey numbers:

• R(3,13): Increased from 60 to 61
• R(3,18): Increased from 99 to 100
• R(4,13): Increased from 138 to 139
• R(4,14): Increased from 147 to 148
• R(4,15): Increased from 158 to 159

These improvements, while seemingly modest, represent years of computational effort compressed into a single meta-algorithm approach.

AlphaEvolve: A New Approach

AlphaEvolve represents a departure from traditional methods in Ramsey theory. Rather than developing bespoke search algorithms for each specific case, this LLM-based code mutation agent generates search algorithms automatically.

This meta-algorithm approach successfully recovered lower bounds for all Ramsey numbers known to be exact and matched the best known lower bounds across many other cases. The researchers note that these include bounds for which previous work does not detail the algorithms used, suggesting AlphaEvolve can discover novel search strategies.

Why This Matters

The significance extends beyond the specific numerical improvements. Virtually all known Ramsey lower bounds are derived computationally, with researchers traditionally developing custom algorithms for each result. AlphaEvolve's ability to produce a single meta-algorithm that works across multiple cases suggests a more scalable approach to tackling these combinatorial challenges.

The work demonstrates how AI techniques can accelerate progress in pure mathematics, potentially opening new avenues for research in areas where computational methods have plateaued. While the improvements here are incremental, the methodology could prove transformative for other unsolved problems in combinatorics and beyond.

Technical Context

The paper, submitted to arXiv on March 10, 2026, and revised the following day, falls under the categories of Combinatorics, Artificial Intelligence, and Computational Complexity. The researchers include Ansh Nagda, Prabhakar Raghavan, and Abhradeep Thakurta from institutions that remain unspecified in the abstract.

For those interested in exploring the technical details, the full paper is available at arXiv:2603.09172, with the current version being v2 from March 11, 2026.