AndroidLife-530 — Android agent benchmark (real phone, real LLM)

AndroidLife runs Android agent tasks against a real phone (via ADB/MobileRun) and a real LLM, and grades the agent on reaching a verifiable device end-state. This repo ships the 530-task corpus plus everything needed to reproduce runs.

Benchmark, or template — your call. The 530 tasks are an extended version of the benchmark, usable as a larger evaluation set for further benchmarking of models beyond the 60-task preview. They are also usable as a template for producing synthetic data: 530 data points are nowhere near enough to train a model, but each record ships its prompt_template, placeholder slots, and vars/ pin files, so the slots can be re-resolved (or regenerated) into many more concrete instances per task without rewriting the corpus.

Formerly published as DrainBench-530 / DailyBench. Canonical dataset id: YuvrajSingh9886/androidlife-530. Legacy YuvrajSingh9886/drainbench-530 redirects here.

Public run trajectories + the 60-task preview live in YuvrajSingh9886/androidlife-public.

Dataset preview (tasks)

Open the Dataset Viewer above (config tasks) for a table of all 530 tasks.

Config	File	What you see
tasks (default)	data/tasks_preview.jsonl	Compact preview columns for browsing
tasks_full	data/AndroidLife_530_v1.jsonl	Full per-task records

Files

File	Content
data/AndroidLife_530_v1.json / .jsonl	The 530-task corpus (Easy 1pt / Medium 3pt / Hard 5pt).
data/tasks_preview.jsonl	Compact rows for the HF Dataset Viewer.
data/tasks_530.md	Human-readable source (canonical prompt text).
data/tasks.md	Wider task list (superset).
data/multiturn_kb_530.json	Knowledge-base profiles for ASK USER - MULTI tasks.
data/ask_user_facts_530.json	Facts for ASK USER SINGLE tasks.
data/hallucination_controls.json	Hallucination-control tasks.
vars/	530-corpus run-time variable values (resolve every [placeholder]).
config/user_config.example	Documented default persona/device config.
fabrication/	Full disclosure of every fabricated test entity.

Task model

• DETERMINISTIC — everything needed is seeded on-device; end state is ADB-verified.
• ASK USER SINGLE — one fact deliberately omitted; agent must ask.
• ASK USER - MULTI — multi-turn dialogue against the KB profile.

Each record carries task_id, exact prompt_text/prompt_template, placeholder slots, difficulty, and grading type. Prompts use [placeholder] slots resolved through the run-time variables in vars/.

Fabricated test data — disclosure

To make deterministic tasks solvable, a controlled fictional persona and fabricated data are seeded on the device. Everything is disclosed honestly in fabrication/ (machine-readable JSON + full human-readable doc). No real phone numbers, personal emails, or real identities appear; all personas are fictional.

License & citation

• Dataset / task content: CC BY 4.0 — reuse allowed with attribution to Yuvraj Singh.
• Harness code: Apache-2.0 in the AndroidLife repo.

If you use AndroidLife — the benchmark, leaderboard, tasks, or results — please credit this work and cite it as:

@misc{singh2026androidlife,
      title={AndroidLife: Real-Phone Android Agent Benchmark for Open-Weight Models and On-Device SLMs},
      author={Yuvraj Singh},
      year={2026},
      howpublished={\url{https://github.com/YuvrajSingh-mist/AndroidLife}},
}

---

AndroidLife — a real-phone Android agent benchmark. 530 tasks, graded on the on-device end state: not a simulator, not a mock API.

Leaderboard + step-by-step replays · All models & datasets · Code

Fuel the next run — a full 60-task suite is ~6.5 h of phone time plus real API spend: