From the source
Microsoft Research's podcast roundup covers new work on coding agents, computer-use agents, workplace AI use, teacher lesson planning and extended-range weather forecasting.
- A self-improving framework called ForkPilot learns when it is worth revisiting earlier decisions and adjusts its strategy to new outcomes, and in tests across six benchmarks and leading AI models it matched the best existing methods while cutting AI processing demands by up to 59%.
- CABRA, a framework for generating coding challenges from scratch, let researchers raise complexity in a controlled way; across thousands of challenges, agents often solved problems with tools but still struggled to understand unfamiliar codebases.
- TeleTune learns reusable skills and workflows from user activity logs and infers the goals behind sequences of actions, and it outperformed existing approaches in evaluations across software and browser tasks.
- An analysis of aggregated Copilot Chat activity found editing documents led classified activities at more than 20 percent, but differences between customer categories grew substantially once the 20 most common activities were set aside.
- A study in low-resource schools in India found Shiksha Copilot cut the time teachers spent on bilingual lesson planning and bureaucracy, with more than a thousand teachers taking part.
