Microsoft Research Focus – Edition #98

Microsoft Research's Focus newsletter, hosted by principal program manager Anson Ho, gathers several recent results from across the company. The edition covers AI translation, a resolved conjecture in high-dimensional geometry, a benchmark and method for keeping coding agents efficient, a new tokenizer tested across 104 languages, and a research intern's account of work at the Tokyo lab. The most immediately actionable item for engineering teams is the coding-agent work, which comes with concrete measurements.
Coding agents write and modify code, but long sessions can exhaust the amount of information a model can process at once. When that happens, agents summarize earlier interactions in a costly step the researchers call 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. The imbalance is striking: a small tail of long-running sessions drives nearly half of spend, which suggests optimization efforts should target exactly those sessions rather than the average one.
To measure the problem, the researchers built CompactionBench, a benchmark for evaluating context compaction during long-running coding tasks. They also proposed APEC, a method that prepares summaries in the background instead of interrupting the agent's work. In their evaluations, APEC improved coding performance while reducing costs by up to 25.4% and execution time by 9.6%. This is my own reading, but the design is intuitive: compaction currently sits on the critical path, so an agent stalls while it summarizes. Moving that work off the critical path trades spare capacity for latency gains, and the reported numbers fit that pattern.
A second thread concerns tokenization. Before models process text they split it into tokens, and how that split is made affects both efficiency and performance, especially across languages. Most tokenizers optimize for compressing text as tightly as possible. A new study argues compression alone is insufficient and introduces the Latent Core Tokenizer, or LCT, which identifies reusable components of words before building a shared vocabulary across languages. Across evaluations spanning 104 languages, LCT produced more linguistically meaningful tokens than leading alternatives while keeping comparable efficiency, and it improved model performance on four multilingual benchmarks. The takeaway the researchers draw is that tokenization quality depends not just on how much text is compressed but on how well language structure survives the process.
On the mathematics side, new research resolves the KLS conjecture, described as one of the central open problems in high-dimensional geometry for the past 31 years. According to the newsletter, the result removes a major theoretical obstacle to understanding the efficiency of randomized algorithms used across science and computing. It establishes a universal principle for log-concave distributions: simple flat cuts can identify their narrowest bottlenecks within a fixed factor, regardless of dimension. The results have been independently validated, and the proof was also formally verified with computer-assisted methods. For context, randomized algorithms underpin sampling and optimization across many scientific workloads, and the conjecture's resolution is a statement about how tightly such distributions can be pinched, which is the kind of structural fact that bounds algorithm behavior. The formal verification is notable on its own, since machine-checked proofs are still uncommon for results of this scale and give the community a stronger basis for building on the work.
In applied AI, Microsoft applied science manager Jinyu Li discusses progress toward communication across languages, what researchers have achieved and what remains difficult, in an episode of the podcast Our Digital Life available on YouTube and podcast platforms. The newsletter frames the work against the long science-fiction dream of a universal translator, citing Star Trek's 1967 version, and treats the goal as closer to reality rather than reached. Speech translation is a demanding setting because errors compound across recognition, translation and synthesis, and latency matters for conversation, though the source does not detail specific systems or results here.
Finally, the edition profiles Lilika Makabe, a research intern at the Microsoft Research lab in Tokyo. Makabe arrived with a computer vision background and worked alongside researchers on robotics and physical intelligence, meaning AI systems that interpret information and act in the physical world. Her account, published on the Microsoft Research blog, emphasizes hands-on experimentation and interdisciplinary collaboration as part of the internship experience.
Why it matters: Teams running coding agents can take a near-term lesson from the CompactionBench study, since long sessions appear to be a disproportionate cost center and background summarization like APEC is a plausible lever, though the reported gains come from the researchers' own evaluations and would need validation on other workloads. Multilingual product teams should watch tokenizer design, where linguistic structure rather than raw compression is the proposed differentiator. The KLS result is foundational rather than shippable, but its formal verification is a signal about how rigorous proof is becoming in high-dimensional geometry.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
What's new in research at Microsoft? Host Anson Ho explores the latest, including AI-powered universal translation, a breakthrough in a 31-year-old mathematics challenge, and advances making AI coding agents faster and more efficient. The post Microsoft Research Focus – Edition #98 appeared first on Microsoft Research .