Benchmark
AndroidControl
Evaluate whether agents can follow instructions and generalize across Android apps and tasks.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
15,283 human demonstrations cover 14,548 unique tasks across 833 Android apps. Each demonstration includes a high-level goal and instructions for individual steps. Authors: Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu and Oriana Riva.
Methodology
Compare agents using high-level goals or detailed step instructions, and distinguish performance within the training domain from generalization to unfamiliar tasks and apps.