AIGenerated
EN · IT · ES

Microsoft researchers solve a 31-year-old math problem and make AI coding agents cut costs by a quarter

A decades-old math puzzle is finally cracked, and a new method helps AI coding tools save money without slowing down.

A still from the announcement video.
Picture: from the announcement video, microsoft.com
How it works

Microsoft researchers cracked a 31-year-old math problem and built methods to make AI coding agents and language tools faster and cheaper

  1. Mathematicians apply new universal rule to geometry problems
  2. AI coding agents run longer without expensive summaries
  3. Tokenizers find word parts shared across 104 languages

What you could do with it

  • A mathematician could apply the new rule about log-concave distributions to unlock progress on other problems in high-dimensional geometry and randomized algorithms 1.
  • A software team running AI coding assistants could keep sessions productive longer and lower their cloud bill by up to about a quarter using the new background summary method 1.
  • A researcher building AI tools for smaller languages could use a tokenizer that finds reusable word parts across 104 languages instead of focusing only on compression 1.

Breaking a decades-old math barrier

For over thirty years, the KLS conjecture stood as one of the central unsolved problems in high-dimensional geometry, a branch of math that studies shapes in spaces with many more directions than the length, width and height we experience daily 1.

Researchers at Microsoft have now resolved it 1. The result establishes a universal rule about log-concave distributions, which are mathematical patterns that appear widely in probability and computing . It shows that simple flat cuts can identify the narrowest bottlenecks inside these patterns within a fixed margin, no matter how many dimensions are involved .

That finding clears away a significant theoretical barrier to grasping how effective randomized algorithms are across science and computing 1. The proof has been independently validated and also verified using computer-assisted methods .

Coding agents that keep their cool for longer

AI coding agents can help people write and modify software, but long working sessions often overwhelm how much information a model can hold at once 1. When that happens, systems summarize earlier interactions through a process called context compaction, which consumes considerable computing resources .

Microsoft researchers examined 13.5 million coding sessions and found that fewer than 8 percent needed this summary step, yet those sessions accounted for 44 percent of total AI processing costs 1. To study the problem systematically, they created CompactionBench, a benchmark for evaluating context compaction during extended coding tasks .

They also introduced APEC, a method that prepares summaries quietly in the background instead of stopping an agent mid-task 1. In tests, APEC boosted coding performance while lowering costs by up to 25.4 percent and trimming execution time by 9.6 percent .

Smarter text-splitting across 104 languages

Before AI models can read or write, they must break text into smaller units called tokens 1. Most systems focus on compressing text as tightly as possible, but a new Microsoft study argues that compression alone leaves value on the table .

Researchers proposed the Latent Core Tokenizer, or LCT, which finds reusable building blocks inside words before constructing a shared vocabulary that works across many languages 1. Across evaluations covering 104 languages, LCT generated more linguistically meaningful tokens than top competing options, with efficiency remaining roughly on par . It also lifted model performance on four multilingual benchmarks .

The team concluded that good tokenization depends not only on how much text can be squeezed together, but on how faithfully the underlying structure of language is preserved 1.

The bigger picture

The edition also pointed listeners toward a separate podcast conversation with Jinyu Li, a manager in Microsoft Applied Science, about how AI is pushing closer to the kind of real-time spoken translation once imagined only in science fiction 1. Meanwhile, intern Lilika Makabe shared how hands-on work at the Microsoft Research lab in Tokyo helped bridge computer vision and physical intelligence, the field behind robots and machines that navigate the real world .

Taken together, the updates sketch a research program that chases speed, clarity and reach — less cost per line of code, tighter mathematical foundations and language tools that stop treating every tongue as an afterthought 1.

How good is it

Microsoft researchers studied 13.5 million coding sessions and found that fewer than 8 percent triggered heavy context compaction, yet those sessions made up 44 percent of total processing costs 1. Their new method cut costs by up to 25.4 percent and execution time by 9.6 percent while also improving coding performance . The 31-year-old math proof was independently validated and verified with computer-assisted methods . For tokenization, tests across 104 languages showed more linguistically meaningful tokens and better model scores on four multilingual benchmarks compared with leading alternatives, though Microsoft did not publish exact rival numbers .

Chart from the announcement: This pie chart compares the small share of coding sessions needing context compaction against their large share of AI processing costs—larger slices mean higher cost impact

Picture: microsoft.com

Key facts

  • The KLS conjecture in high-dimensional geometry remained open for 31 years; Microsoft researchers say it is now resolved, with independent validation and computer-assisted proof.
  • An analysis of 13.5 million coding sessions found fewer than 8 percent needed costly context compaction, but those sessions drove 44 percent of processing costs.
  • APEC, a background-summary method from the same Microsoft work, cut execution time by 9.6 percent and costs by up to 25.4 percent in coding-agent tests.
  • The Latent Core Tokenizer (LCT) was tested across 104 languages, producing more linguistically meaningful tokens and improving four multilingual benchmarks.

Sources

  1. Microsoft Research Focus – Edition #98 - Microsoft Research

Watch

Pressing the box loads the video from YouTube, which then receives your visit under its own privacy policy. Until then nothing is requested from YouTube.

If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work

Keep reading