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)→
More from coding & agent
- Are Gemini 4 Neon fans actually using it in Antigravity? Coder asks — iScienceLuvr · 2026-11-22
- '2,000 lines from Claude, 10-second Codex review': devs fret over AI engineering quality — _jaydeepkarale · 2026-11-22
- Cairnmind: open-source MCP server enforcing read-before-write for agent memory — Zestyclose_Bus_8031 · 2026-11-22
- Team releases Part 2 of $102,725 AI episode, sharing 50,000-character Claude skill — azed_ai · 2026-11-22
- 20GB open model builds flight simulator offline on a plane in minutes — heyshrutimishra · 2026-11-22
- Entire GTA 6 Apple Maps Site Built by GPT-6.2, Dev Says It Nails Tasks First Try — Dimillian · 2026-11-22