Microsoft Research Focus Edition 97: ForkPilot cuts agent processing by 59%, BioEmu2 enters public preview

• ForkPilot, a self-improving agentic framework, matches the performance of leading methods across six benchmarks while reducing AI processing demands by up to 59%.
• BioEmu2, a foundation model for understanding biomolecular structure, dynamics, and function, is now available in public preview via the Microsoft Research GitHub repository.
• CABRA, a framework for generating coding challenges, found that understanding unfamiliar codebases remains a primary bottleneck for coding agents, even when they can use tools to solve problems.
• Probabilistic Bias Correction (PBC) won a global competition for extended-range weather forecasting and is being deployed in Kenya to improve monsoon forecasts for farmers.

Agent Frameworks

• TeleTune learns reusable skills and workflows from user activity logs, outperforming existing approaches in software and browser task evaluations.
• ForkPilot adapts its strategy based on new outcomes to determine when it is worth revisiting earlier decisions.

Research Findings

• Analysis of Copilot Chat activity showed that editing documents accounted for more than 20 percent of classified activities, but differences between customer categories became pronounced when the 20 most common activities were excluded.
• Shiksha Copilot, a human-in-the-loop AI tool, reduced time spent on bilingual lesson planning and bureaucracy for more than a thousand teachers in low-resource schools in India.