From the source
Microsoft Research's newsletter roundup covers a proof resolving the 31-year-old KLS conjecture in high-dimensional geometry, which the company says removes a major theoretical obstacle to understanding randomized algorithms.
- A study of 13.5 million coding sessions found fewer than 8% needed context compaction, yet those sessions accounted for 44% of total AI processing costs.
- To address that, researchers built COMPACTIONBENCH, a benchmark for evaluating context compaction in long-running coding tasks, and APEC, a method that prepares summaries in the background instead of interrupting an agent's work.
- In evaluations, APEC improved coding performance while cutting costs by up to 25.4% and execution time by 9.6%.
- Separately, researchers introduced the Latent Core Tokenizer, which identifies reusable components of words before building a shared vocabulary across languages; across 104 languages it produced more linguistically meaningful tokens than leading alternatives while keeping comparable efficiency, and improved model performance on four multilingual benchmarks.
- The edition also features a conversation on AI speech translation and a Tokyo research intern's move from computer vision to physical intelligence.






