01
Zhang et al. · 2023
Why?
Start here for the agent itself: an LLM multimodal agent that taps and swipes like a person, learns apps by exploring or watching demos, and never needs backend access.
Loading…
One moment while we fetch this page.
From what a phone agent is, through GUI grounding, to AndroidWorld and the harder realism / robustness benchmarks that followed.
The full collection in the community’s reading order.
01
Zhang et al. · 2023
Why?
Start here for the agent itself: an LLM multimodal agent that taps and swipes like a person, learns apps by exploring or watching demos, and never needs backend access.
02
Cheng et al. · 2024
Why?
Names the capability most eval failures hide behind: GUI grounding (finding the right screen element from an instruction), and introduces ScreenSpot.
03
Rawles et al. · 2024
Why?
The workhorse eval harness: a live Android environment with 116 parameterized tasks, reward signals, and init/success/teardown logic so runs stay reproducible. Everything harder on this shelf is reacting to this baseline.
04
Xu et al. · 2024
Why?
Pairs a reproducible Android env (138 tasks / 9 apps) with open-model training results, so you see both how to score agents and how far open LLMs/LMMs still lag before the harder suites.
05
Kong et al. · 2025
Why?
The ‘AndroidWorld is saturating’ follow-up: longer-horizon, multi-app, vague-instruction tasks (≈2× steps, far more cross-app) plus interactive / MCP-style tool use. Read once the clean harness feels too easy.
06
Sui et al. · 2026
Why?
Closed-source daily apps with no internal state to peek at: 350 tasks / 94 apps, scored with GRADE. Use after the open harnesses to see the opacity gap in real phone workflows.
07
Liu et al. · 2026
Why?
Robustness as a missing eval axis: perturbs AndroidWorld along state/transition/action so clean benchmark scores stop lying. Ends the shelf with ‘how far from deployment’ rather than another task list.
For
Builders who want a short path into evaluating Android GUI agents.
Covers
AppAgent, SeeClick grounding, AndroidWorld / AndroidLab / MobileWorld as eval harnesses, then AndroidDaily closed-source apps and AndroidReality robustness — the evaluation and training methodologies that form the basis for Android app agents.
Leaves out
Desktop/web GUI agents, or deep training recipes beyond what AndroidLab already covers for eval context.