Microsoft Research Focus #98 covers KLS proof, agent compaction, tokenizer

Microsoft Research Focus #98 covers KLS proof, agent compaction, tokenizer

Edition #98 of Microsoft Research Focus is a short podcast, based on the Research Focus newsletter and hosted by Anson Ho. It runs through five items from across Microsoft, with chapter marks from 0:35 to 3:57, and points listeners to longer write-ups on the Microsoft Research blog.

First comes a people story. At the Microsoft Research lab in Tokyo, research intern Lilika Makabe joined with a background in computer vision and worked alongside researchers exploring robotics and physical intelligence, which the episode defines as AI systems that can interpret information and act in the physical world. The episode presents her experience as a sign that internships offer more than technical training: hands-on experimentation, interdisciplinary collaboration and everyday conversations with colleagues from different backgrounds widened her research perspective while she helped shape the lab's growing community.

Second is AI speech translation. The episode notes that in 1967 Star Trek introduced the fictional Universal Translator, and says advances in AI are now bringing that possibility closer to reality. In a conversation on the podcast Our Digital Life, Microsoft Applied Science Manager Jinyu Li explores how AI is advancing communication across languages, the progress researchers have made, and the challenges that remain. The episode is on YouTube and podcast platforms.

Third is mathematics. For 31 years the KLS conjecture has been one of the central open problems in high-dimensional geometry. The episode says new research has resolved it, removing a major theoretical obstacle to understanding the efficiency of randomized algorithms used across science and computing. The result establishes a universal principle for log-concave distributions: simple flat cuts can identify their narrowest bottlenecks within a fixed factor, regardless of dimension. According to the episode, the results have been independently validated and the proof has also been formally verified using computer-assisted methods.

Fourth is coding agents. Long sessions can max out the amount of information a model can process at once, so agents summarize earlier interactions in a resource-intensive step called context compaction. A new study examined 13.5 million coding sessions and found that fewer than 8% required compaction, yet those sessions accounted for 44% of total AI processing costs. Researchers built COMPACTIONBENCH, a benchmark for evaluating context compaction during long-running coding tasks, and introduced APEC, a method that prepares summaries in the background instead of interrupting the agent's work. In evaluations APEC improved coding performance while reducing costs by up to 25.4% and execution time by 9.6%.

Fifth is tokenization. Models split text into tokens before processing it, and how the text is divided affects efficiency and performance, especially across languages. Most tokenizers aim to compress text as efficiently as possible, but a new study suggests compression alone is not enough. Researchers introduced the Latent Core Tokenizer (LCT), which identifies reusable components of words before building a shared vocabulary across languages. In evaluations spanning 104 languages, LCT produced more linguistically meaningful tokens than leading alternatives while keeping comparable efficiency across languages, and it improved model performance across four multilingual benchmarks. The findings suggest better tokenization depends not only on how much text can be compressed but on how well the underlying structure of language is preserved.

Key facts

  • The KLS conjecture, open for 31 years in high-dimensional geometry, is described as resolved: simple flat cuts find the narrowest bottlenecks of log-concave distributions within a fixed factor, regardless of dimension. The proof was independently validated and formally verified with computer-assisted methods.
  • A study of 13.5 million coding sessions found fewer than 8% needed context compaction, but those sessions made up 44% of total AI processing costs.
  • COMPACTIONBENCH benchmarks context compaction in long coding tasks; APEC prepares summaries in the background and cut costs by up to 25.4% and execution time by 9.6% in evaluations while improving coding performance.
  • The Latent Core Tokenizer (LCT) builds a shared vocabulary from reusable word components; across 104 languages it gave more linguistically meaningful tokens than leading alternatives at comparable efficiency, and improved results on four multilingual benchmarks.
  • The episode also covers Tokyo research intern Lilika Makabe and a conversation with Applied Science Manager Jinyu Li on AI speech translation.

Why it matters

The episode bundles three results that touch core machinery. A resolved KLS conjecture is presented as removing a major theoretical obstacle to understanding the efficiency of randomized algorithms. The compaction study points at where coding-agent money goes: a small share of sessions (under 8%) carries 44% of processing costs. The tokenizer work argues that better tokenization depends on preserving the structure of language, not only on compressing text.

Who it affects

Mathematicians and computer scientists working with randomized algorithms and log-concave distributions are the audience for the KLS result. Teams building or running coding agents are the audience for CompactionBench and APEC. People building multilingual models are the audience for LCT. The intern and speech translation segments are aimed at a general listener interested in Microsoft Research's culture and in cross-language communication.

How to use it

The episode is a quick overview, with chapters at 0:35 (internship), 1:28 (language barrier), 2:10 (KLS), 2:56 (coding agents) and 3:57 (tokenization). Each item points to a longer piece: the related links include write-ups on the intern's journey, AI speech translation, the KLS conjecture, CompactionBench and the Latent Core Tokenizer, and the Jinyu Li conversation is in the Our Digital Life podcast on YouTube and podcast platforms.

How solid is it

This is a first-party roundup narrated by the host, not a paper, and no direct quotations from researchers appear in it. For the KLS result the episode says the results were independently validated and the proof was formally verified with computer-assisted methods, but it does not say how or by whom. The size of the 'fixed factor' is not given. For APEC, no baseline is given for the 25.4% and 9.6% reductions, and 'up to' applies to the cost figure only. The four LCT benchmarks and the 'leading alternatives' are not named, and no numeric LCT scores are given.

Risks and caveats

Headline numbers come from a short summary, so details that decide how far they generalize are missing: the source does not say which coding agents, models or session data sets were behind the 13.5 million sessions. No authors or institutions are named for the KLS result, the CompactionBench and APEC study, or the LCT study. What Jinyu Li concluded in the Our Digital Life conversation is not stated beyond the fact that he explores progress and remaining challenges. The LCT study is described as suggesting, not proving, that compression alone is not enough.

“But a new study suggests that compression alone isn’t enough.”

— Anson Ho, host, Microsoft Research Focus