Microsoft Researcher Proves 31-Year-Old Conjecture With AI

Two researchers, including Xinzhi Zhang of the Machine Learning and Optimization group at Microsoft Research, have proved the Kannan–Lovász–Simonovits (KLS) conjecture, a problem in high-dimensional geometry first posed in 1995, Microsoft said in a research blog post on 21 October 2026. The blog describes “a human-directed agentic research process in which AI systems explored mathematical possibilities while researchers evaluated, pruned, and redirected the search.”

The post says “Roughly one hundred research trajectories were explored” before the decisive idea appeared. It also says that in a separate track the authors made repeated unguided attempts with several leading AI models, “using roughly the same total token budget as the main process,” and that “That effort fell short of the goal.”

Business Pill · MULTI-AGENT SYSTEMS

A short Business Engineer explainer on multi-agent systems: one agent or a team of them. It teaches the general idea only and says nothing about Microsoft, the researchers or the proof in this piece.

What Was Proved

In the blog’s words, KLS “asks whether, for a broad class of high-dimensional probability distributions, the best flat or hyperplane cut is always nearly as effective as the best cut of any shape.” Put simply, it asks whether the easiest way to split a high-dimensional shape is always close to the best way, no matter how many dimensions the shape has.

The question matters outside geometry. “Because the geometry of these cuts governs how quickly random walks can explore a distribution, KLS also plays a central role in the theory of high-dimensional sampling,” the post says. The paper’s introduction adds that bounds on the KLS constant control the complexity of sampling, volume computation and log-concave integration, and underlie algorithms for convex optimization.

The proof is in a paper titled An O(1) Bound for the KLS Constant, by Zhao Song, listed as an independent researcher, and Xinzhi Zhang of Microsoft Research. It was first posted to arXiv on 1 October 2026; the version dated 4 October completes the proof. Before it, the best known bound, from July 2026, still grew slowly with the dimension, the post says.

Selected rows from Table 1 of arXiv 2610.01447 (Song and Zhang, version of 4 October 2026), restated in words. Lower growth is better; the conjecture asked for a bound that does not grow with the dimension at all.

What Was Posted on X

AI commentator Rohan Paul summarised the result on X on 22 October. His post says AI agents explored roughly 100 research trajectories while a researcher cut weak directions between rounds, and that the recursive method and the error bound that finished the proof are credited to the two researchers. He also describes the conjecture as a random walker that no narrow passage can trap; the piece relies on Microsoft’s post for the facts.

Who Did What: the Researchers and the Agents

The post is specific about the division of labour. “The recursive method, its repeated application, and the proof that the accumulated errors remain bounded are original contributions of the researchers,” it says. “AI tools helped explore candidate approaches, carry out detailed derivations, prepare the manuscript, and perform internal proof checks. The researchers determined the mathematical direction, selected the promising ideas, assembled them into the final argument, and verified the proof.”

The search itself ran in rounds. “As multiple agents explored variations of an idea in parallel, the researcher examined their intermediate result, eliminated unpromising directions, and passed the most promising result to the next round,” the post says. Working that way, the authors “explored more than 100 candidate approaches, many involving lengthy calculations with higher-order tensors.”

The post draws its own conclusion from the unguided comparison: “The eventual success depended on both the capabilities of the AI systems and the mathematical judgment of the researchers.”

How the Result Was Checked

• A parallel proof. On 4 October, Pierre Bizeul, Boaz Klartag and Joseph Lehec posted “Presenting a proof of the Kannan-Lovasz-Simonovits conjecture” to arXiv. Its abstract says: “The crucial step in the proof is the very recent criterion from Song and Zhang involving high-order derivatives of tilt averages.”
• A third proof. The post also describes a third proof which, it says, “appeared separately on the same day” and “relies on a closely related idea.” Separately, a paper by Krishnakumar Balasubramanian and Shiva Kasiviswanathan, first posted to arXiv on 6 October, gives a proof with an explicit Poincaré bound of 25, according to its abstract.
• Machine checking. The post says “a Lean 4 formalization” was subsequently released, “combining the key criterion from the first Microsoft paper with additional research,” and that the authors have “shortened their proof and verified the streamlined argument in Lean,” with plans to release both “after final review.”

Connecting the Coverage

This is the latest in a run of AI-assisted mathematics we have covered. Our piece on AlphaProof Nexus reported on 13 October that Google DeepMind’s system, which has every step checked in the Lean proof language, solved 9 of the 353 open Erdős problems it attempted, according to a paper in Science. Our OpenAI math piece covered the company’s 6 October release of 722 manuscripts from an internal model, results that OpenAI said it was publishing on GitHub.

The KLS post adds a different data point: a comparison, inside one project, between human-directed agents and unguided model runs on a similar token budget. It sits next to the Google Cloud and UIUC paper in our harness piece, which reported 77.41% against 61.62% on a task benchmark with the same model and different harnesses. Our Recursive piece covered another lab saying its own automated research system helped build a model, and our Microsoft stack piece quoted Satya Nadella calling the company’s stack “model-forward and agent-first.”

What It Means for Businesses

As we read it, the useful lesson is about process, not about mathematics. The post describes a loop in which people choose directions and cut losers while agents do the long derivations, and it credits the result to that loop rather than to a single model run.

As we read it, the same token budget producing different outcomes depending on how the work is directed is the cost-to-outcome question that buyers of AI agents face in ordinary workflows.

The Structural Read

As we read it, the post credits the result to a loop: agents generate and work through candidate approaches in parallel, and a researcher prunes them between rounds.

As we read it, the post is careful about credit. It names the recursive method and the bounded-error argument as the researchers’ own contributions and lists the AI’s jobs as exploration, derivations, manuscript preparation and internal proof checks.

As we read it, the checking is what separates this from a claim: a parallel proof by other mathematicians that calls the Song and Zhang criterion “the crucial step in the proof,” and a Lean formalization, per the post.

Three Implications

FOR COMPANIES DEPLOYING AGENTS As we read it, the post’s comparison suggests that how a token budget is directed can matter as much as its size; the post gives no budget figure, so the size of that effect is not established.

FOR LARGE AI VENDORS As we read it, the post shows Microsoft presenting AI as a research instrument under human direction, the same framing as Microsoft’s stack described as “model-forward and agent-first” in Nadella’s letter.

FOR RESEARCH ORGANISATIONS As we read it, the checking stack the post describes, independent proofs plus a machine-checked formalization, is the part that turns AI-assisted output into an accepted result.

The Business Engineer Lens

This story maps onto the Business Engineer framework AI & The Harness Theory.

The framework’s core line: “The next professional jump is not a better prompt. It is an operating system that wraps your AI — a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”

As we read it, the KLS project is a harness in that sense: researchers’ judgment, rounds of pruning and the models’ raw capability, combined in one process. The framework was published in April 2026 and does not discuss this proof.

What Is Not Established

The blog post and the paper abstract we read do not name the AI models or agents used, or the size of the token budget. The post does not say whether the paper has been accepted by a peer-reviewed journal; its confirmation rests on the independent proofs and the Lean formalization it describes. We read the blog post, the arXiv abstracts of the three papers and the introduction of the Song and Zhang paper, not the full 140-page proof, and we have not checked the mathematics.

The Bottom Line

Two researchers, including Microsoft Research’s Xinzhi Zhang, proved the 31-year-old KLS conjecture using what Microsoft calls human-directed AI. The researchers set the direction and verified the proof, the post says, after AI agents explored roughly one hundred research trajectories, while unguided attempts on roughly the same token budget fell short.