Microsoft Research Focus – Edition #97

Microsoft Research published the 97th edition of its Research Focus series on October 29, 2026, with program manager Sowbhagya Suresh presenting seven recent projects spanning AI agents, workplace AI use, education, weather and biology.

Three items concern agents. ForkPilot is a self-improving framework that learns when revisiting earlier decisions is worthwhile; Microsoft says it matched the best existing methods across six benchmarks while cutting AI processing demands by up to 59%. CABRA generates coding challenges from scratch, and across thousands of them researchers found that agents handled many problems with tools but struggled to understand unfamiliar codebases. TeleTune learns reusable skills from user activity logs and infers the goals behind actions, and reportedly beat existing approaches on software and browser tasks.

An analysis of aggregated Copilot Chat activity found document editing was the top classified use, at more than 20 percent, but differences between customer categories grew substantially once the 20 most common activities were set aside. In a study involving more than a thousand teachers at low-resource schools in India, Shiksha Copilot reduced time spent on bilingual lesson planning and bureaucracy, with many participants reporting lower stress.

Microsoft also highlighted Probabilistic Bias Correction, a machine-learning method that learns from a weather model's past errors to improve two-to-six-week forecasts. It won a global forecasting competition and is being used for monsoon forecasts for farmers in Kenya. Finally, BioEmu2, a foundation model for biomolecular structure, dynamics and function, is now in public preview on GitHub, which the team says could eventually support the development of more effective medicines.