Analysis: The Mathematical Aptitudes of LLMs

Written from the perspective of Gowers's Weblog

To ensure this post remains useful for future readers (perhaps a month from now), I must establish the context: I am writing this shortly after OpenAI's announcement that they have cracked ten significant problems in theoretical computer science and mathematics. Among these were the first construction of a non-sofic group and a proof regarding the multicolour Ramsey number $R(3,3,...,3)$, demonstrating that it grows superexponentially in $k$.

The first achievement represents a milestone in group theory; the second was a major open question in Ramsey theory that I did not expect to see solved in my lifetime now realize can be solved by AI.

Note on Timing: I am discussing LLM capabilities with the understanding that they are evolving at a breakneck pace. By the time you read this, these observations may serve primarily as a historical snapshot of the landscape in early August 2026.

The Current State of AI in Mathematics

While these ten results are staggering, it does not appear that LLMs have completely eclipsed human mathematicians across the board. If they had, their massive computational speed would have triggered a deluge of new theorems far greater than what we are seeing.

This leads to a compelling question: What specific types of mathematical problems are LLMs actually good at, and where do they still struggle?

While I lack a "crisp" classification system, we can attempt to refine the answer by eliminating unlikely theories and identifying patterns in the evidence.

The "Counterexample" Hypothesis

One might notice that LLMs are capable of both proving difficult theorems and finding counterexamples. However, a pattern emerges: the most high-profile "wins" have predominantly been counterexamples.

If we want to argue that LLMs are uniquely suited for finding counterexamples, we must complete the following checklist:

• Define precisely what constitutes "finding a counterexample."
• Provide a theoretical explanation for why LLMs excel at this specific task.

The Nuance of Counterexamples

Defining a "counterexample" is more complex than it seems. A naive definition might look like this:

Statement: "Every object of type X has property P." $\rightarrow$ Counterexample: "Here is an object of type X that lacks property P."

However, consider Vinogradov's Theorem, which posits that every sufficiently large positive integer is the sum of three primes. If we negate this, we get:

$\forall N \in \mathbb{Z}^+, \exists n \ge N$ such that $n$ is not a sum of three primes.

In a strictly logical sense, Vinogradov proved that the negation is false by showing that for all $n \ge N$, the property holds. If we viewed the original theorem as a search for a counterexample to the negation, then every great theorem would be a "counterexample." This is clearly the wrong way to categorize it.

The Role of Quantifiers

The difference lies in the "interesting" variable. In Vinogradov's work, the focus was not on searching for a specific $n$, but on the universal quantifier $\forall n \ge N$.

Many formal results involve an alternation of quantifiers ($\forall, \exists, \forall \dots$). The crux of the problem is identifying which quantifier is the "driver" of the discovery.

Case Study: Banach-Mazur Compactum

To illustrate this, let's look at finite-dimensional normed spaces.

Suppose we have two $k$-dimensional normed spaces, $X$ and $Y$, and a linear map $T: X \to Y$. We call $T$ a $\lambda$-isomorphism if there exists a scaling such that for all $x \in X$: $\|Tx\| \le \lambda \|x\|$

The Banach-Mazur distance $d(X, Y)$ is the minimum $\lambda$ for which such a map exists. This allows us to treat the set of isometry classes of $k$-dimensional normed spaces as a metric space (via the logarithm of the distance), known as the Banach-Mazur compactum.

A pivotal result by Fritz John states:

Every $k$-dimensional space $X$ has a distance $d(X, \ell_2^k) \le \sqrt{k}$.

The logic of this proof involves finding an ellipsoid of maximal volume $\mathcal{E}$ inside the unit ball of $X$.

Visualizing the Components

Below are the mathematical elements involved in these types of discussions:

In this context, the "search" is for the optimal ellipsoid, which mirrors the way one might search for a counterexample, yet the result is a universal bound. This distinction is where the real investigation into LLM "intuition" must begin.