# AndroidBench — Research references

Reviewed 2026-09-18. These are design references, not benchmarks reproduced in this device pilot.

| Primary source | Design implication |
|---|---|
| [MobileWorld](https://github.com/Tongyi-MAI/MobileWorld) | Separate interaction, capability channels and environment restoration. |
| [AndroidLab](https://github.com/THUDM/Android-Lab) | Verify real system state independently from UI appearance. |
| [MobileBench-OL](https://github.com/xiaomi-research/mobilebench-ol) | Long tasks can exist inside a single app; count obligations and recovery. |
| [AmbiBench](https://arxiv.org/html/2602.11750v1) | Evaluate necessary clarification without rewarding redundant questions. |
| [KnowU](https://arxiv.org/html/2604.08455v1) | Distinguish explicit tasks, synthetic preferences and proactive behavior. |
| [Mobile-Bench-v2](https://arxiv.org/abs/2505.11891) | Use replaceable slots and accept multiple valid paths. |
| [AndroidControl](https://github.com/google-research/google-research/blob/master/android_control/README.md) | Step prediction and autonomous task success require separate evaluation. |
| [AndroidLife](https://github.com/YuvrajSingh-mist/AndroidLife) | Inspect public tasks and trace provenance; public subsets are not additional tasks. |
| [MobilePA-Bench](https://arxiv.org/html/2608.23035v1) | Distinguish tool planning from physical phone GUI evaluation. |
| [GMA](https://arxiv.org/html/2608.27477v1) | State tracking and retained context require controlled ablations. |
| [AndroidReality](https://arxiv.org/html/2608.07775v1) | Record perturbation type, trigger and recovery obligations. |
| [MobileMem](https://arxiv.org/html/2608.13606v1) | Long-term memory needs a temporal history, not a single-session claim. |

Candidate contribution: evaluate execution qualification, user corrections, artifact provenance, recovery side effects and duplicate execution together on real devices. This remains a research direction, not a demonstrated novelty claim or a published-paper result.

Website information architecture reference: [PostTrainBench](https://posttrainbench.com/). AndroidBench uses its own data and visual implementation and has no affiliation with the reference project.