Human-directed AI research helps establish a dimension-free bound for the KLS conjecture

A problem in high-dimensional geometry asks how badly a simple straight cut can compare with the best possible boundary through a probability distribution. For 31 years, the Kannan–Lovász–Simonovits conjecture predicted that the gap could be bounded by a fixed constant, however many dimensions the problem involved. October papers now establish that bound, following research in which mathematicians directed AI agents through a large set of possible approaches.[1][2][3]

Zhao Song and Microsoft's Xinzhi Zhang present their result in ‘An O(1) Bound for the KLS Constant’. A Microsoft Research account published on 21 October describes how their exploration produced a new criterion and then a proof. A separate team—Pierre Bizeul, Boaz Klartag and Joseph Lehec—also posted a proof, explicitly identifying the Song–Zhang criterion as a crucial step.[1][2][3]

Why dimension independence matters

The conjecture concerns log-concave probability distributions, a family important in geometry, statistics and optimisation. A useful informal picture is a region with probability mass on each side of a narrow passage. How difficult it is to cross those bottlenecks affects how efficiently a random walk can explore the distribution and generate samples.[1][3]

Earlier bounds became worse as the number of dimensions increased, even as researchers progressively slowed that growth. An O(1) result gives a universal bound independent of dimension. The statement does not promise a tiny numerical constant or an immediate product speedup; it strengthens the theoretical guarantee available to an entire family of sampling and optimisation arguments.[1][2]

The chronology separates progress from completion

An iterated logarithm grows extraordinarily slowly, but it still grows. Microsoft's account explains that the final step applied the recursive improvement again and controlled the errors it accumulated. That change is what turned a stronger dimension-dependent estimate into the conjectured universal statement.[2]

How the researchers used AI

Microsoft reports that the researchers explored more than 100 candidate approaches with AI assistance. Agents helped with derivations, calculations, variants and checking, while the mathematicians pruned unsuccessful directions and decided which structures deserved further work. The useful output was a criterion that could be placed inside a rigorous argument.[2]

The account also reports unsuccessful attempts to solve the problem without the same human steering under a comparable reasoning-token budget. That comparison supports the importance of the workflow in this case. It is a more specific research finding than a claim that a model can routinely solve arbitrary open problems.[2]

Several routes to scrutiny

The independent paper by Bizeul, Klartag and Lehec makes the connection explicit in its abstract: their proof uses the recent criterion involving higher-order derivatives of tilt averages. Its presence gives mathematicians another complete argument to examine and places the result within a wider research effort, rather than a single company's account of an AI achievement.[3]

Microsoft also describes formal-verification work around the criterion and related results, alongside plans for a streamlined formal version. That is distinct from asserting that every line of the original manuscript has already been mechanically checked as one complete formal proof.[2]

The mathematical gain is the dimension-free bound itself. The research-process gain is a documented example of computational exploration becoming useful through expert selection and proof construction. Its eventual consequences for practical algorithms will come through the work that builds on that firmer theoretical foundation.[1][2][3]

Report an error or suggest a correction