Benchmark
Android in the Zoo (AitZ)
Evaluate smartphone GUI agents using screen observations, action decisions and intermediate reasoning.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
18,643 screen-action pairs with Chain-of-Action-Thought annotations support smartphone task evaluation and agent training.
Methodology
Agents use screen observations and action context to predict the next smartphone action. The original project supplies evaluation code; the paper studies context modeling and training with its annotations.