Microsoft Research finds 8% of coding-agent sessions drive 44% of AI costs, plus a KLS proof

This edition of Microsoft Research Focus, hosted by Anson Ho, includes a study of 13.5 million coding sessions. Fewer than 8% needed context compaction, the process of summarizing earlier work when a session runs out of room, yet those sessions accounted for 44% of total AI processing costs. That is a striking concentration of spend in a small slice of usage.

Microsoft researchers built COMPACTIONBENCH to evaluate how agents handle compaction, and a method called APEC that prepares summaries in the background instead of stopping the agent. They report cost reductions of up to 25.4% and execution time down 9.6%, with improved coding performance. These are Microsoft’s numbers and need replication, but the direction is right. Long sessions are where teams get the most value and where bills climb, so cheaper long tasks are what make agents usable for real engineering. The previous edition had the same theme of cheaper, more honest agents (Edition #97), as does Microsoft’s Agent Lightning framework for training agents in the harness they ship in (coverage).

Microsoft says new research has resolved the KLS conjecture, a central open problem in high-dimensional geometry for 31 years, with independent validation and computer-assisted formal verification. In plain terms, it shows that simple flat cuts can find the narrowest bottlenecks of a broad class of distributions within a fixed factor, regardless of dimension. That matters for understanding how efficiently randomized algorithms run. The roundup doesn’t say what role, if any, AI played, so none should be assumed. It lands amid other claimed breakthroughs on long-open problems, where verification and access have become the debate (OpenAI’s math claims).

The Latent Core Tokenizer identifies reusable word parts before building a shared vocabulary. Across 104 languages it produced more linguistically meaningful tokens at comparable efficiency and improved results on four multilingual benchmarks. Tokenizers built mainly to compress text tend to serve some languages better than others, so preserving linguistic structure is a quiet step toward models that work well beyond English.

The translation segment is more aspiration than result: a pointer to a podcast conversation with Applied Science Manager Jinyu Li about progress and remaining challenges toward a universal translator. It’s a worthy goal, and one where errors in medical, legal or diplomatic settings carry real harm, so evidence should come before excitement. The internship segment profiles a computer vision researcher who worked with robotics and physical-intelligence researchers at the Tokyo lab. Each item links to a fuller write-up on the Microsoft Research blog. More in our Microsoft coverage and research coverage.