# Microsoft Research Focus #98: Geometry, efficient coding agents and multilingual AI

- Source: MechNext — https://mechnext.eu/en/article/70-microsoft-research-focus-98-geometry-efficient-coding-agents-and-multilingual-ai
- Published: 2026-11-13
- Updated: 2026-11-13
- Category: Industry News
- Language: en (Markdown in other languages: ?lang=de, ?lang=sk, ?lang=pl, ?lang=zh, ?lang=ru)

> The edition brings together a resolution of the 31-year-old KLS conjecture, context-compaction research, a multilingual tokenizer, and discussions of robotics and cross-language communication.

**Microsoft Research Focus – Edition #98**, hosted by Anson Ho, covers research from across Microsoft. Its subjects range from robotics and physical intelligence to AI-enabled communication, high-dimensional geometry, long-running coding agents, and multilingual tokenization.

### From computer vision to physical intelligence

Research intern Lilika Makabe joined the Microsoft Research lab in Tokyo with a background in computer vision. There, she worked with researchers exploring robotics and physical intelligence. This part of the edition situates her experience at the meeting point between prior work in visual computing and a research environment concerned with robotics and physical intelligence.

A second discussion turns to communication across languages. Microsoft Applied Science Manager Jinyu Li discusses communication enabled by AI, the progress researchers have made, and the challenges that remain. The account therefore presents both sides of the current research picture: advances already achieved and work that has not yet been resolved.

### A 31-year problem in high-dimensional geometry

The edition’s mathematical result is new research resolving the KLS conjecture. The conjecture is described as a central open problem in high-dimensional geometry that had remained open for 31 years. The breakthrough establishes that simple flat cuts can identify the narrowest bottlenecks of log-concave distributions within a fixed factor, regardless of dimension.

The report also specifies how the result was checked. The KLS findings were independently validated, and the proof was formally verified through computer-assisted methods. Those are distinct forms of scrutiny: independent validation of the results and formal verification of the proof. Both accompany the description of the mathematical breakthrough.

### The cost concentration in long coding sessions

The work on coding agents draws on a study of 13.5 million coding sessions. Fewer than 8% of those sessions required context compaction, yet that subset accounted for 44% of total AI processing costs. The comparison matters because it distinguishes frequency from cost: compaction was needed in only a small share of sessions, while those sessions represented a much larger share of processing expenditure.

Researchers developed **COMPACTIONBENCH** to evaluate context compaction during long-running coding tasks. They also introduced **APEC**, a method that prepares summaries in the background instead of interrupting a coding agent’s work. One contribution supplies a way to evaluate compaction; the other changes how summary preparation is incorporated into an ongoing coding task.

In evaluations, APEC improved coding performance while reducing costs by up to 25.4% and execution time by 9.6%. The reported outcome therefore combines three dimensions from those evaluations: coding performance, processing cost, and execution time. The figures describe the evaluated method rather than every possible coding session.

### Reusable word components across 104 languages

The multilingual work addresses how text is divided into tokens before model processing. Researchers introduced the **Latent Core Tokenizer**, which identifies reusable word components before constructing a shared vocabulary across languages. The ordering is central to the method as described: reusable components are found first, and the shared multilingual vocabulary is built afterward.

Evaluations covered 104 languages. Across them, the Latent Core Tokenizer produced more linguistically meaningful tokens than leading alternatives, maintained comparable efficiency across languages, and improved model performance on four multilingual benchmarks. The reported evidence consequently spans token quality, cross-language efficiency, and model results rather than relying on a single comparison.

Taken together, Edition #98 is a collection of clearly separated research developments. Makabe’s internship provides the robotics and physical-intelligence perspective. Li’s discussion covers progress and remaining challenges in AI-enabled communication. The KLS work contributes a precise geometric result backed by independent validation and formal verification. The coding-agent studies quantify a concentrated cost problem and assess a background-summary method. Finally, the tokenizer evaluations extend across 104 languages and four multilingual benchmarks.

The common thread is not that these projects solve the same problem, but that the edition states a concrete research question or result for each area. Readers receive specific scales and outcomes: 31 years for the KLS conjecture, 13.5 million coding sessions in the compaction study, cost and time changes for APEC, and 104 languages in the tokenizer evaluations.
