Microsoft Unveils ForkPilot, Cutting AI Workload by Up to 59%

In Research Focus issue 97, released on October 29 local time, Microsoft presented a batch of AI research spanning everything from agents that handle multi-step tasks on their own to weather forecasting and protein research. Among them, ForkPilot reportedly matched the performance of strong existing methods while cutting the amount of work the AI has to process by up to 59%.

When does AI decide to try again?

ForkPilot is a research framework that lets AI learn when to revisit earlier decisions and try a different approach. When new results come in, it changes strategy too. Microsoft evaluated ForkPilot with leading AI models across six benchmarks. The idea is to reduce the resources spent on repeated attempts, improving the operational efficiency of AI entrusted with complex tasks.

In CABRA, a study that evaluates coding AI, understanding unfamiliar code emerged as one of the toughest challenges. CABRA generates new coding tasks and adjusts their difficulty to probe where AI falls short. The researchers observed AI solving coding problems with tools across thousands of tasks. The models still struggled to grasp the structure and contents of codebases they had not seen before.

Microsoft also introduced TeleTune, aimed at AI that gets work done through clicks and keystrokes. It learns reusable task procedures from records of people using software, then infers the purpose behind those actions. Microsoft said it outperformed existing methods on software and browser task benchmarks. The distinguishing feature is its use of human work logs as training material for AI.

How is AI used at work and in classrooms?

Looking only at the most common use cases, the way people use AI at work appears much the same everywhere. When researchers aggregated and analyzed Copilot Chat activity across several customer segments, document editing was the single largest category at more than 20% of classified activity. But once the 20 most common activities were set aside, differences between customer segments became far more pronounced. The researchers stressed that understanding an organization’s AI needs requires looking at what employees actually do and the outcomes they need, not just usage volume.

In India, the team studied Shiksha Copilot, a tool whose output teachers review and use themselves. More than 1,000 teachers at schools with limited educational resources took part. The study reported that the tool reduced the time spent preparing bilingual lessons and handling administrative work. Teachers put the time they saved into classroom activities that boost student engagement, and many participants reported lower stress. The teachers adapted AI-generated materials to fit their own classrooms.

Can we predict the weather two to six weeks out?

On the weather front, Microsoft introduced PBC, which improves forecasts two to six weeks ahead. It learns from patterns in weather models’ past errors and corrects for them. Microsoft said PBC posted the top score in an international competition on improving forecasts over that time range. It has also been deployed in Kenya, where it is being used to improve monsoon forecasts that matter to farmers, the company explained.

BioEmu2, which studies the structure, motion, and function of biomolecules such as proteins, arrived as a public preview. It was developed by the AI for Science team at Microsoft Research, building on the existing Biomolecular Emulator. The company explained that a better understanding of protein function could help in developing effective drugs.

The research presented here focuses on lightening AI’s workload and applying it to human work and scientific research. For coding AI, the open challenge is understanding existing code; for organizational AI adoption, it is pinning down specific job tasks. How much time these improvements actually save in real work is something to watch.