Curated reading list: evaluating mobile-use agents on Android, from AppAgent to AndroidReality

East-Muffin-6472 · reddit · 2026-11-22

The author surveyed the literature on evaluating Android mobile GUI agents to build their own everyday-task benchmark (AndroidLife), curating a papercase reading-order shelf. Training+eval methodology: AppAgent, SeeClick (GUI grounding), AndroidWorld, AndroidLab. Evaluation-only benchmarks: MobileWorld (MCP-augmented), AndroidDaily (real closed-source apps), AndroidReality (realism gap). Goal: verifiable evals of everyday tasks with models running locally on the phone; feedback welcome.

Related event: Curated Reading List for Android Mobile Agent Evaluation(2 posts)→

Original post →

More from coding & agent

coding & agent channel →