A 322M typed-decision policy for Android GUI control. Accessibility tree only · no vision · no autoregressive generation.
| Parameters | AndroidControl-High Type | AndroidControl-High Grounding | Latency (RTX 4090, bs1) |
|---|---|---|---|
| 322M | 76.7 | 61.7 | ~40 ms |
Every step of a recorded validation episode (chosen by a fixed rule) with Laya-Android's predictions: blue = predicted element, green = gold. The model never sees the screenshots; it reads the accessibility tree. Gold history, not a live run. An episode with a failure: docs/results.md.
Given a goal, the last three actions and the actionable UI elements of the Android accessibility tree, Laya-Android picks the next operation and the target element in one forward pass. It is a fine-tune of Laya by Convai Innovations, trained only on AndroidControl.
Same task form as Laya-Android: typed choice questions answered with probabilities. Both models received byte-identical requests (state + operation and target questions) on the same 1,000 AndroidControl-High test steps.
| Laya-Android (322M, local) | Jev 1.13.0 (TypeSafe API) | |
|---|---|---|
| Type | 77.2 | 67.2 |
| Grounding | 62.8 | 56.7 |
| Operation ECE (lower is better) | 0.037 | 0.066 |
| Operation Brier (lower is better) | 0.338 | 0.487 |
| Target ECE (lower is better) | 0.061 | 0.140 |
| Median latency | 39 ms (RTX 4090) | 236 ms (API round trip) |
Reliability diagram: when a model says it is 80% sure, it should be right 80% of the time (the diagonal). Jev 1.13.0 was called through the TypeSafe API on 2026-10-07; it is a general decision model, not trained on AndroidControl. Jev is not fully deterministic: an earlier run with the same requests gave Type 67.6, Grounding 56.4 and operation ECE 0.055 (raw: results/baselines/).
Same 1,000 test steps, same input (goal, last 3 actions, the same accessibility candidates), same InfiGUI-R1 judge, one RTX 4090 at batch 1. The LLMs are used zero-shot (not trained on AndroidControl), thinking disabled, and asked for a JSON action; vLLM runs them in bf16, Ollama in 4-bit GGUF. The gap in latency comes from selecting instead of generating; the gap in accuracy also reflects that only Laya-Android was trained on this dataset.
All 16 models
| Model | Params | Serving | Type | Grounding | Median latency | p95 | vs Laya-Android |
|---|---|---|---|---|---|---|---|
| Laya-Android | 322M | Laya-Android (local) | 77.2 | 62.8 | 39 ms | 48 ms | 1x |
| Qwen3.8 27B | 27B (Q4_K_M) | Ollama GGUF | 76.8 | 64.5 | 575 ms | 694 ms | 14.9x slower |
| Qwen3.5 4.2B | 4.2B (Q4_K_M) | Ollama GGUF | 67.3 | 52.8 | 193 ms | 252 ms | 5.0x slower |
| jev-1.13.0 | undisclosed | TypeSafe API | 67.2 | 56.7 | 236 ms | 295 ms | 6.1x slower |
| Qwen3.5 4B | 4B (bf16) | vLLM bf16 | 67.1 | 50.5 | 242 ms | 305 ms | 6.3x slower |
| Gemma 4 25B | 25B (Q4_K_M) | Ollama GGUF | 64.8 | 56.6 | 201 ms | 256 ms | 5.2x slower |
| Qwen3.5 9.7B | 9.7B (Q4_K_M) | Ollama GGUF | 63.6 | 51.6 | 229 ms | 273 ms | 5.9x slower |
| Gemma 4 E2B | E2B (bf16) | vLLM bf16 | 58.4 | 23.6 | 155 ms | 210 ms | 4.0x slower |
| Qwen3.5 2B | 2B (bf16) | vLLM bf16 | 57.5 | 23.0 | 149 ms | 202 ms | 3.9x slower |
| Gemma 4 8.0B | 8.0B (Q4_K_M) | Ollama GGUF | 57.4 | 45.0 | 171 ms | 202 ms | 4.4x slower |
| Qwen3.5 1.9B | 1.9B (Q8_0) | Ollama GGUF | 57.4 | 24.5 | 126 ms | 144 ms | 3.3x slower |
| Gemma 4 E4B | 7.5B (Q4_K_M) | Ollama GGUF | 57.1 | 44.1 | 146 ms | 175 ms | 3.8x slower |
| Gemma 4 E2B | 4.6B (Q4_K_M) | Ollama GGUF | 56.9 | 22.7 | 116 ms | 139 ms | 3.0x slower |
| Ministral 3 14B | 14B (Q4_K_M) | Ollama GGUF | 53.2 | 42.0 | 285 ms | 424 ms | 7.4x slower |
| Qwen3.5 0.8B | 0.8B (Q8_0) | Ollama GGUF | 41.3 | 0.0 | 124 ms | 158 ms | 3.2x slower |
| Qwen3.5 0.8B | 0.8B (bf16) | vLLM bf16 | 40.9 | 0.4 | 125 ms | 178 ms | 3.2x slower |
AndroidControl-High, 8,444 test steps / 1,543 episodes, public InfiGUI-R1 evaluator, evaluated once with the checkpoint frozen on validation (protocol):
| Model | Params | Type | Grounding | Full SR |
|---|---|---|---|---|
| Laya base (zero-shot) | 322M | 29.0 | 11.8 | N/A |
| Laya-Android | 322M | 76.7 | 61.7 | N/A |
| Δ | +47.6 | +49.9 |
Full SR is N/A: it requires generated text and app names, which v0 does not produce. Screenshot-based VLMs on the same evaluator report higher numbers (e.g. InfiGUI-R1-3B: 82.7 Type / 74.4 Grounding) but solve a different task (pixel coordinates instead of choosing among accessibility candidates). Per-split results including unseen apps, tasks and categories: docs/results.md.
| Hardware | Precision | Median | p95 | Peak VRAM |
|---|---|---|---|---|
| RTX 4090 | bf16 autocast | 39.7 ms | 43.5 ms | 1.66 GB |
Latency stays near 40 ms up to 50 candidates and 1,024 input tokens. Batch sweep and length scaling: docs/benchmark.md.
git clone https://github.com/Byun11/laya-android && cd laya-android
pip install -r requirements.txt
python examples/quickstart.py # loads ByunByun/laya-androidimport laya
from serialization import OP_DESC, Q_OP, Q_TARGET # src/serialization.py
agent = laya.load("ByunByun/laya-android")
state = {"goal": goal, "history": last_3_actions, "ui": ["[%d] %s" % (i, c) for i, c in enumerate(candidates)]}
questions = {
"operation": {"type": "choice", "instructions": Q_OP, "criteria": OP_DESC},
"target": {"type": "choice", "instructions": Q_TARGET, "criteria": {str(i): c for i, c in enumerate(candidates)}},
}
out = agent.system_one(state, questions, max_len=3072, head_max_len=1536)
out["answers"]["operation"]["choice"], out["answers"]["target"]["choice"] # ('CLICK', '4')Build candidates with src/candidates.py; other label formats are out of distribution.
- No visual input: unlabeled icons and custom-drawn UI are hard to tell apart.
- No text or app-name generation, so no Full Step SR; AndroidControl-Low not evaluated.
- Policy Joint drops 9–12 points on unseen apps, tasks and categories.
- Weak on rare actions (WAIT, BACK, horizontal scroll, LONG_PRESS).
- Offline, step-level benchmark with gold history; not an autonomous task success rate.
| docs/results.md | per-split results, strict grounding, operation recall, examples, training progression, related work |
| docs/benchmark.md | same-input comparison details, latency by candidates, input length and batch size |
| docs/training.md | model diagram, data, candidate extraction, input format, actions, training, reproduction |
| docs/eval_protocol.md | evaluation protocol, frozen before scoring |
@software{laya_android_2026,
title = {Laya-Android},
author = {Byun, Jaeyeon},
year = {2026},
version = {0.3.0},
url = {https://github.com/Byun11/laya-android}
}Built on Laya (Convai Innovations, Apache-2.0; not affiliated), mmBERT-base (MIT) and AndroidControl (Google Research, Apache-2.0). Evaluator: InfiGUI-R1 (Apache-2.0). Licensed Apache-2.0, see LICENSE and NOTICE.