Seven researchers published a detailed case study of using an AI system to tighten the best known bounds on the Grothendieck constant, and the most useful part of the paper is its self-criticism. The system was good at technical execution and had some steps judged novel by domain experts. It was weak at research judgment and at keeping an accurate picture of where the work stood. That combination is the honest description of what AI mathematical collaboration currently is.
Key facts
- Submitted August 11, 2026 by Alan Li, Rahul Saha, Anton Xue, Swarat Chaudhuri, Adam Klivans, Pravesh Kothari, and Raghu Meka, per the paper.
- The team tightened the known bounds on the Grothendieck constant to between 6 pi over 11 and pi over twice the natural log of one plus root two, minus a small margin.
- The system combined a reasoning model, a coding agent, file-based memory, calibrated internal verification, and asynchronous human steering.
- The reported weaknesses are research judgment and maintaining an accurate research state.
What the constant is
The Grothendieck constant, introduced in 1953, captures how much you lose when you replace a hard combinatorial optimization problem with the continuous relaxation used to approximate it. That gap is not an abstraction. It sets limits on how well whole families of approximation algorithms can possibly perform, which is why theoretical computer scientists care about it as much as analysts do.
Its exact value has been unknown for seventy years. Progress happens by squeezing the upper and lower bounds closer together, in increments, over decades. That makes it an unusually good test case for AI-assisted research: the target is precisely defined, the improvements are verifiable, and nobody can hand-wave a result past the referees.
The framing that makes this paper useful
Most AI-for-mathematics coverage answers the question "did the machine do something impressive." This paper answers a more useful one: how do you actually use these systems, and where do they let you down.
The setup they describe has four parts. A reasoning model does the mathematical thinking. A coding agent runs computations and checks. File-based memory holds the state of the research across sessions, because no context window is large enough to carry a months-long project. And calibrated internal verification means the system checks its own claims and knows roughly how confident to be, rather than asserting everything with equal certainty.
The last piece is the human. The authors call it asynchronous steering, and the word choice is deliberate. The human is not reviewing every step, and is not absent either. They intervene periodically to redirect, which is exactly the shape you would expect given the failure mode the paper reports.
The failure the paper names
The system was weak at research judgment, meaning deciding which of several plausible directions is worth the next week of effort. And it was weak at maintaining an accurate research state, meaning knowing what had already been tried, what had already been ruled out, and how the current attempt relates to the overall plan.
Those two failures are the same failure. Research judgment is largely a function of accurate state: you cannot tell which direction is promising if you have lost track of which directions already failed and why. This is the identical problem documented in this week's study of agent instruction files that triple in size because nobody remembers why a rule exists. Rationale decays faster than the artifact it justifies, and once it is gone the system keeps working but stops steering.
File-based memory is the mitigation these authors reached for, and it clearly was not sufficient on its own, or the human steering would not have been necessary. Our note on agent memory covers why writing things down solves less of this than people expect: a file records what happened, not what it meant.
Read alongside the other math result this week
The contrast with Anthropic's zeta function result is instructive, because the two runs failed in opposite directions and both worked anyway. Anthropic's model burned through 650 ideas before its subagent swarm found the productive one, with a non-mathematician human contributing little beyond encouragement. That is brute-force exploration compensating for exactly the judgment deficit this paper measures. Here, seven working mathematicians supplied the judgment and used the system for execution.
Both produced real results. Neither describes a system that knows what to work on. The scaling question the field has not answered is whether judgment improves with model capability or whether it is a distinct thing that has to be engineered, and this paper is a data point for the second reading.
The caveat
This is a case study by the authors of their own work, and the strongest claim in it, that experts judged some AI-originated steps novel, is the authors judging steps produced by a system they built and operated. That is not independent verification, and the paper does not pretend otherwise. The bounds themselves are checkable mathematics and stand on their own. The methodological conclusions are one team's experience, valuable precisely because they are specific and unflattering, but they are one team's experience. What would settle it is other groups running the same setup on unrelated problems and reporting whether the same two weaknesses show up. Related reading: measuring AI by task length and test-time compute.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)