Microsoft Research Focus – Edition #98

This edition of Microsoft Research Focus, hosted by Anson Ho, covers a Tokyo lab internship, AI translation, the resolution of a 31-year-old math problem, cheaper coding agents, and a tokenizer tested across 104 languages. Most of it is a sampler. Two items deserve more attention than a roundup format gives them.

The first is 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.

We think the coding-agent finding is the most practical item here. Long agent sessions are exactly where teams get the most value, and also where the bills climb. Microsoft researchers built a benchmark, COMPACTIONBENCH, to evaluate how well agents handle compaction, and a method called APEC that prepares summaries in the background instead of stopping the agent to do it. They report cost reductions of up to 25.4% and execution time down 9.6%, with improved coding performance. Those are Microsoft's own numbers, and we'd want to see them replicated, but the direction is right. Making agents cheaper on long tasks is what makes them usable for real engineering work, not just demos.

The math result is the headline-grabber. 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 we won't assume one.

The tokenizer work deserves a nod for fairness reasons. The Latent Core Tokenizer identifies reusable word parts before building a shared vocabulary, and across 104 languages it produced more linguistically meaningful tokens at comparable efficiency, while improving results on four multilingual benchmarks. Tokenizers built mainly to compress text tend to serve some languages better than others. Work that preserves 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 the progress and remaining challenges. A universal translator is a worthy goal, and also one where errors in medical, legal or diplomatic settings carry real harm. We'll want evidence, not science fiction comparisons, before getting excited.

Research Focus is Microsoft Research's short-form roundup based on its newsletter. Each item links to a fuller write-up on the Microsoft Research blog. The internship segment profiles a computer vision researcher who worked alongside robotics and physical-intelligence researchers at the Tokyo lab.