When a Quarter of Math Papers Admit a Machine Helped

What if the most important number in AI-for-mathematics right now isn’t a benchmark score at all, but a disclosure rate?

Between March 1 and August 20, 2026, researchers collected all 32,944 submissions to the arXiv Mathematics category and looked at a single question: how many of them said out loud that an AI system was involved. In March, 4.75% did. By August, 24.14% did. That’s a fivefold jump in under six months, in a field whose publication culture is famously conservative about credit.

I care about that curve more than I care about the latest headline result, and I want to explain why from an architecture perspective.

Proof construction is the interesting column

Buried in the same data are two counts that deserve more attention than they’re getting. Among the disclosing submissions, 1,712 involved at least one substantive mathematical contribution from an AI system. And 1,225 papers reported AI involvement in proof construction specifically.

That distinction matters. Using a model to clean up LaTeX, rephrase an abstract, or sanity-check a reference is a text task. Participating in proof construction is something else: it requires maintaining a goal stack, proposing intermediate lemmas, recognizing when a branch is dead, and backing out of it without losing the thread. Those are agent properties, not language properties. You don’t get them for free from scaling next-token prediction, and you don’t get them from a single forward pass.

So when 1,225 papers report that kind of involvement, what we’re actually observing is the deployment footprint of long-horizon reasoning systems inside a domain that has an unusually honest verifier built into it. Mathematics is one of the few places where the reward signal is close to ground truth. A proof either closes or it doesn’t. For anyone building agents, that makes math less a showcase and more a laboratory.

The disclosure rate is a measurement instrument, not a capability curve

I want to be careful here, because it’s easy to read 4.75% → 24.14% as “AI got five times better at math in five months.” It doesn’t say that. Disclosure rates are shaped by at least three forces tangled together: actual tool use, changing norms about what must be reported, and social permission to admit it without looking like you cheated.

My read is that the third force moved fastest. The tools were already in people’s workflows; what changed is that saying so stopped being professionally risky. If that’s right, then the 4.75% figure from March was always an undercount, and the real adoption curve is flatter and started earlier than the headline suggests.

This is a recurring problem in how we measure agent deployment. We don’t have instrumentation on the work itself, so we infer from self-report, and self-report tracks norms as much as behavior. Anyone drawing capability conclusions from disclosure data should hold them loosely.

Hundreds of results, almost no specifics

Then there’s the other half of the story. A new internal model at an unnamed company has reportedly produced a broad range of new mathematical results, including a solution to the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis. The sources describing this don’t provide a definitive account of what the advances actually are or how they were obtained.

From where I sit, that gap is the real problem, and it’s bigger than any individual claim. For an agent researcher, a result without a trace is nearly useless. I can’t tell you whether that system searched, planned, called a formal verifier, ran a critic loop, or decomposed goals. I can’t tell you how many attempts preceded the successful one, which is the single most informative number for understanding an agent’s architecture. A system that finds a proof in one shot and a system that finds it in ten thousand are different machines doing different things, even when the output is identical.

The arXiv study, for all its limits, gives us a population-level view we can argue with. The internal-model announcements give us outcomes we can only accept or doubt. Those are different epistemic objects, and the field is currently treating them as the same kind of evidence.

What I’d actually want logged

If the goal is understanding, not announcement, the useful disclosures look like this:

• Attempt counts and compute spent per result, not just the successful trajectory
• Whether a formal proof assistant was in the loop, and at which step
• The division of labor between human and system, lemma by lemma
• Failure cases on neighboring problems, which tell you where the capability boundary sits

A workshop premise circulating recently takes superhuman mathematical ability as a starting assumption rather than a debate. Fine as a thought experiment. But assuming the destination doesn’t tell us anything about the vehicle, and the vehicle is what the rest of us have to build.

One in four math papers now says a machine helped. The next useful question is which machine, doing what, and how many times it failed first.

🕒 Published: October 8, 2026