Microsoft Research: 31-year-old KLS conjecture solved, CompactionBench cuts agent costs

Microsoft Research · rss · 2026-11-13

Microsoft Research Focus #98 rounds up recent work from across Microsoft:

KLS conjecture resolved: New research settles the 31-year-old central open problem in high-dimensional geometry, establishing a universal principle for log-concave distributions—simple flat cuts find their narrowest bottlenecks within a fixed factor regardless of dimension. Independently validated and formally verified with computer-assisted methods.
Coding-agent context compaction: A study of 13.5M coding sessions found fewer than 8% required compaction, yet those accounted for 44% of total AI processing costs. The team released CompactionBench and introduced APEC, which prepares summaries in the background, improving coding performance while cutting costs up to 25.4% and execution time 9.6%.
Multilingual tokenization: The Latent Core Tokenizer (LCT) identifies reusable word components before building a shared vocabulary; across 104 languages it yields more linguistically meaningful tokens and improves four multilingual benchmarks.
AI speech translation: Applied science manager Jinyu Li discusses progress toward a universal translator.
Also: a research intern's path from computer vision to physical intelligence at MSR Asia–Tokyo.

More from Research