ForkPilot cuts AI processing up to 59% across six benchmarks

This edition of Microsoft Research Focus bundles several agent-related efforts. ForkPilot targets a specific weakness: agents that need to adjust when things go wrong may consume a lot of resources with no guarantee of a better result. The framework learns when revisiting earlier decisions is worth it. Separately, the CABRA framework generates coding challenges from scratch with controlled complexity. Across thousands of such challenges, agents often succeeded by using tools but struggled to understand unfamiliar codebases, pointing to code comprehension rather than code modification as a key obstacle. TeleTune addresses a related gap for computer-use agents, which click and type to complete tasks. Activity logs usually lack the intent behind each action and mix tasks together; TeleTune learns reusable skills from those logs and infers the goals behind action sequences, outperforming existing approaches in software and browser evaluations. On the usage side, an analysis of aggregated Copilot Chat activity found document editing led all classified activities at more than 20 percent, while differences between customer categories became substantially more pronounced once the 20 most common activities were set aside. Elsewhere, a study of Shiksha Copilot in low-resource Indian schools, with more than a thousand teachers participating, found reduced time on bilingual lesson planning and bureaucracy. And Probabilistic Bias Correction took top honors in a global extended-range forecasting competition and is being deployed in Kenya for monsoon forecasts.