Microsoft Research’s 97th Research Focus edition highlights several projects involving AI agents, workplace use, education, weather forecasting and biomolecular science.

The episode discusses ForkPilot, a framework that learns when to revisit earlier decisions during long-horizon agent tasks. Microsoft says tests across six benchmarks and leading AI models showed performance comparable to existing methods while reducing processing demands by up to 59%.

Another project, CABRA, generates coding challenges with adjustable complexity. According to Microsoft’s researchers, evaluations across thousands of challenges found that coding agents often used tools successfully but continued to struggle with understanding unfamiliar codebases. TeleTune is designed to learn reusable skills and workflows from user activity logs, including inferred goals behind sequences of actions. Microsoft says it outperformed existing approaches in software and browser-task evaluations.

The edition also summarizes an analysis of aggregated Copilot Chat activity. Editing documents accounted for more than 20% of classified activities, while less common activities showed greater differences between customer categories. Microsoft says common usage statistics may therefore overlook more specialized organizational needs.

Microsoft reports that a study involving more than 1,000 teachers in low-resource schools in India found Shiksha Copilot reduced time spent on bilingual lesson planning and administrative work. The episode also covers Probabilistic Bias Correction, a framework for forecasts two to six weeks ahead. Microsoft says its BioEmu2 biomolecular foundation model is available in public preview.