Microsoft Research Focus #98: KLS Conjecture Solved and Coding Agent Costs Cut by 25.4%

• Microsoft researchers resolved the KLS conjecture, a 31-year-old open problem in high-dimensional geometry, with a proof formally verified using computer-assisted methods.
• A new study analyzing 13.5 million coding sessions found that fewer than 8% required context compaction, yet those sessions accounted for 44% of total AI processing costs.
• Researchers introduced APEC, a background summarization method that improved coding performance while reducing costs by up to 25.4% and execution time by 9.6%.
• The Latent Core Tokenizer (LCT) was developed to create more linguistically meaningful tokens across 104 languages, improving model performance on four multilingual benchmarks.

Research Highlights

• The KLS conjecture resolution establishes a universal principle for log-concave distributions, removing a theoretical obstacle to understanding randomized algorithm efficiency.
• APEC prepares summaries in the background to avoid interrupting agent work, addressing the resource-intensive nature of context compaction in long-running tasks.
• LCT identifies reusable word components before building a shared vocabulary, suggesting that preserving language structure is as important as compression efficiency.