-
Notifications
You must be signed in to change notification settings - Fork 2
Expand file tree
/
Copy pathindex.html
More file actions
497 lines (446 loc) · 21.9 KB
/
Copy pathindex.html
File metadata and controls
497 lines (446 loc) · 21.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="title" content="AndroidControl-Pro: Revealing the True Potential of GUI Agents through Benchmark Purification">
<meta name="description" content="We reveal systemic flaws in existing GUI benchmarks and introduce AndroidControl-Pro, a purified benchmark for accurate evaluation. Our new SOTA model, Magma-R1, trained on just 2,400 high-quality samples, proves that data quality fundamentally trumps quantity in training high-performance GUI agents.">
<meta name="keywords" content="GUI Agents, Benchmark Purification, Data Quality, Reinforcement Learning, Autonomous Agents, AndroidControl, Magma-R1, Deep Learning, Computer Vision, AI">
<meta name="author" content="Anonymous">
<meta name="robots" content="index, follow">
<meta name="language" content="English">
<!-- Performance optimization -->
<meta name="theme-color" content="#2563eb">
<meta name="msapplication-TileColor" content="#2563eb">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="default">
<!-- Open Graph -->
<meta property="og:type" content="article">
<meta property="og:site_name" content="Anonymous Research">
<meta property="og:title" content="AndroidControl-Pro: Revealing the True Potential of GUI Agents through Benchmark Purification">
<meta property="og:description" content="Our work introduces a purified benchmark, AndroidControl-Pro, to address systemic flaws in GUI agent evaluation, and a new SOTA model, Magma-R1, demonstrating that data quality is more critical than quantity.">
<meta property="og:url" content="https://YOUR_DOMAIN.com/AndroidControl-Pro">
<meta property="og:image" content="static/images/data_quality_vs_quantity.png">
<meta property="og:image:width" content="1200">
<meta property="og:image:height" content="630">
<meta property="og:image:alt" content="Data Quality vs Quantity in GUI Agent Training">
<meta property="article:published_time" content="2024-09-01T00:00:00.000Z">
<meta property="article:author" content="Anonymous">
<meta property="article:section" content="Research">
<!-- Twitter -->
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:site" content="@YOUR_TWITTER_HANDLE">
<meta name="twitter:creator" content="@feizuo_handle">
<meta name="twitter:title" content="AndroidControl-Pro: Revealing the True Potential of GUI Agents through Benchmark Purification">
<meta name="twitter:description" content="We show that data quality, not quantity, is the key to high-performance GUI agents with our new benchmark AndroidControl-Pro and model Magma-R1.">
<meta name="twitter:image" content="static/images/data_quality_vs_quantity.png">
<meta name="twitter:image:alt" content="Data Quality vs Quantity in GUI Agent Training">
<!-- Academic citations -->
<meta name="citation_title" content="AndroidControl-Pro: Revealing the True Potential of GUI Agents through Benchmark Purification">
<meta name="citation_author" content="Anonymous">
<meta name="citation_publication_date" content="2024">
<meta name="citation_conference_title" content="Under Review">
<meta name="citation_pdf_url" content="static/images/AndroidControl-Pro-Magma-Manuscript.pdf">
<title>AndroidControl-Pro - Anonymous | Research</title>
<link rel="icon" type="image/x-icon" href="BA_icon.jpg">
<link rel="apple-touch-icon" href="BA_icon.jpg">
<!-- Preconnect for performance -->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link rel="preconnect" href="https://cdn.jsdelivr.net">
<link rel="preconnect" href="https://cdnjs.cloudflare.com">
<!-- Critical CSS - Load synchronously -->
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bulma@0.9.4/css/bulma.min.css">
<!-- Non-critical CSS - Load asynchronously -->
<link rel="preload" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.1.1/css/all.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<!-- Fallback for browsers that don't support preload -->
<noscript>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.1.1/css/all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
</noscript>
<style>
body {
font-family: 'Inter', sans-serif;
}
.publication-title {
font-weight: 800;
margin-bottom: 1rem !important;
line-height: 1.1 !important; /* 调整行高让两行更紧凑 */
}
.publication-authors {
margin-bottom: 0.5rem !important;
}
.hero.teaser {
background-color: #fafafa !important;
border: none !important;
margin-top: 0 !important;
}
.hero.teaser .subtitle {
margin-top: 1rem !important;
color: #4a4a4a !important;
font-size: 1.1rem !important;
line-height: 1.4 !important;
}
#main-method-image {
max-width: 90% !important;
border-radius: 8px;
box-shadow: 0 4px 12px rgba(0,0,0,0.1);
margin: 0 auto;
display: block;
}
/* 调整按钮样式，参考第一张图 */
.publication-links .button {
margin: 0.25rem;
font-size: 0.9rem;
}
.section.contributions ul {
list-style-type: '✨ ';
padding-left: 2em;
}
.section.contributions li {
padding-bottom: 0.5em;
}
.bibtex-header {
display: flex;
align-items: center;
justify-content: space-between;
}
.copy-bibtex-btn {
background-color: #f5f5f5;
border: 1px solid #dbdbdb;
border-radius: 4px;
padding: 5px 10px;
cursor: pointer;
transition: all 0.3s ease;
}
.copy-bibtex-btn:hover {
background-color: #e8e8e8;
}
.copy-bibtex-btn.copied {
background-color: #48c78e;
color: white;
}
/* Scroll to top button */
.scroll-to-top {
position: fixed;
bottom: 20px;
right: 20px;
background: #2563eb;
color: white;
border: none;
border-radius: 50%;
width: 50px;
height: 50px;
cursor: pointer;
opacity: 0;
transition: opacity 0.3s ease;
z-index: 1000;
box-shadow: 0 2px 10px rgba(0,0,0,0.2);
}
.scroll-to-top.visible {
opacity: 1;
}
.scroll-to-top:hover {
background: #1d4ed8;
transform: translateY(-2px);
}
/* 减少section之间的间距 */
.section {
padding: 2rem 1.5rem !important;
}
/* 让整体布局更紧凑 */
.container.is-max-desktop {
max-width: 1152px !important;
}
</style>
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap" rel="stylesheet">
<!-- Structured Data for Academic Papers -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "ScholarlyArticle",
"headline": "AndroidControl-Pro: Revealing the True Potential of GUI Agents through Benchmark Purification",
"description": "We reveal systemic flaws in existing GUI benchmarks and introduce AndroidControl-Pro, a purified benchmark for accurate evaluation.",
"author": [{
"@type": "Person",
"name": "Anonymous"
}],
"datePublished": "2024-09-01",
"publisher": {
"@type": "Organization",
"name": "Under Review"
},
"url": "https://YOUR_DOMAIN.com/AndroidControl-Pro",
"image": "data_quality_vs_quantity.png",
"keywords": ["GUI Agents", "Benchmark Purification", "Data Quality", "Reinforcement Learning", "Autonomous Agents"],
"abstract": "On-device virtual assistants like Siri and Google Assistant are increasingly pivotal, yet their capabilities are hamstrung by a reliance on rigid, developer-dependent APIs. GUI agents offer a powerful, API-independent alternative, but their adoption is hindered by the perception of poor performance...",
"isAccessibleForFree": true,
"about": [{
"@type": "Thing",
"name": "GUI Agents"
}, {
"@type": "Thing",
"name": "Benchmark Purification"
}, {
"@type": "Thing",
"name": "Data Quality"
}]
}
</script>
</head>
<body>
<!-- Scroll to Top Button -->
<button class="scroll-to-top" onclick="scrollToTop()" title="Scroll to top" aria-label="Scroll to top">
<i class="fas fa-chevron-up"></i>
</button>
<main id="main-content">
<section class="hero" style="padding-bottom: 1rem;">
<div class="hero-body" style="padding-bottom: 1rem;">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">
AndroidControl-Pro:<br>
Revealing the True Potential of GUI Agents through Benchmark Purification
</h1>
<div class="is-size-5 publication-authors" style="margin-bottom: 0.5rem;">
<span class="author-block">
Anonymous Authors
</span>
</div>
<div class="is-size-5 publication-authors" style="margin-bottom: 1rem;">
<span class="author-block">Under Review</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="static/images/AndroidControl-Pro-Magma-Manuscript.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/YOUR_REPO_HERE" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/your-dataset" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-database"></i>
</span>
<span>Dataset</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/your-model" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-robot"></i>
</span>
<span>Model</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/YOUR_ARXIV_ID" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser" style="padding-top: 1rem; padding-bottom: 2rem; background-color: #fafafa; border: none;">
<div class="container is-max-desktop">
<div class="hero-body" style="padding-top: 1rem; padding-bottom: 1rem;">
<img src="static/images/method.png" alt="Overall pipeline of our work" id="main-method-image" style="max-width: 90%; margin: 0 auto; display: block;"/>
<h2 class="subtitle has-text-centered" style="margin-top: 1rem; font-size: 1.1rem; color: #4a4a4a;">
<b>Our Integrated Pipeline.</b> Top: The Magma-R1 training process with GRPO. Bottom: The systematic benchmark purification process, transforming the flawed AndroidControl into the reliable AndroidControl-Pro.
</h2>
</div>
</div>
</section>
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
On-device virtual assistants like Siri and Google Assistant are increasingly pivotal, yet their capabilities are hamstrung by a reliance on rigid, developer-dependent APIs. GUI agents offer a powerful, API-independent alternative, but their adoption is hindered by the perception of poor performance, as 3B parameter models score as low as 60% on benchmarks like AndroidControl, far from viability for real-world use. Our work fundamentally challenges this perception by identifying the root cause: the benchmark itself is flawed. During post-training experiments, we discovered that the AndroidControl benchmark contains a significant rate of ambiguities and factual errors, leading to a severe underestimation of agent capabilities. To address this, we introduce AndroidControl-Pro, a new benchmark created through a systematic purification pipeline. On this reliable benchmark, state-of-the-art models achieve success rates approaching 80% on challenging tasks, reframing on-device GUI agents as a near-term reality. Furthermore, we demonstrate that data quality is paramount: our new SOTA model, Magma-R1, trained on just 2,400 curated samples, decisively outperforms models trained on over 31k noisy data points. We release our benchmark and model to accelerate the development of the next generation of powerful, on-device virtual assistants.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section contributions">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Key Contributions</h2>
<div class="content has-text-left">
<ul>
<li>We identify and systematically analyze critical flaws in a mainstream GUI agent benchmark, revealing that poor benchmark quality is a key bottleneck hindering the perceived viability of on-device GUI agents.</li>
<li>We propose a reproducible, semi-automated benchmark purification pipeline to address these flaws, providing a methodology for creating more reliable evaluation tools.</li>
<li>We release <b>AndroidControl-Pro</b>, a rigorously purified new benchmark. On this benchmark, we show that existing models are far more capable than previously believed, achieving performance that makes on-device GUI agents a practical possibility.</li>
<li>We open-source <b>Magma-R1</b>, a new state-of-the-art model that, through training on a small, high-quality dataset, proves that <b>data quality is the most critical factor</b> for unlocking high performance in GUI agents.</li>
</ul>
</div>
</div>
</div>
</div>
</section>
<section class="hero is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Key Results & Insights</h2>
<!-- 第一个结果 - 数据质量 vs 数量 -->
<div class="columns is-centered is-vcentered" style="margin-bottom: 3rem;">
<div class="column is-6">
<img src="static/images/data_quality_vs_quantity.png" alt="Data Quality vs. Quantity Graph" style="width: 100%; border-radius: 8px; box-shadow: 0 4px 12px rgba(0,0,0,0.1);"/>
</div>
<div class="column is-6">
<h3 class="title is-4">Quality over Quantity</h3>
<p class="content">
Our Magma-R1 model, trained on just <strong>2,400 purified samples</strong>, decisively outperforms models trained on over <strong>31k noisy samples</strong>. This demonstrates that data quality is fundamentally more important than data quantity in training high-performance GUI agents.
</p>
</div>
</div>
<!-- 第二个结果 - 系统性基准缺陷 -->
<div class="columns is-centered is-vcentered" style="margin-bottom: 3rem;">
<div class="column is-6">
<h3 class="title is-4">Systematic Benchmark Flaws</h3>
<p class="content">
We identified <strong>three major error types</strong> in the original AndroidControl benchmark: Unclear Tasks, Multiple Valid Actions, and Wrong Ground Truth. These systematic flaws led to severe underestimation of agent capabilities.
</p>
</div>
<div class="column is-6">
<img src="static/images/casestudy.png" alt="Examples of benchmark deficiencies" style="width: 100%; border-radius: 8px; box-shadow: 0 4px 12px rgba(0,0,0,0.1);"/>
</div>
</div>
<!-- 第三个结果 - 评估方法改进 -->
<div class="columns is-centered is-vcentered">
<div class="column is-6">
<img src="static/images/sub_method.png" alt="Grounding Optimization from Point to Box" style="width: 100%; border-radius: 8px; box-shadow: 0 4px 12px rgba(0,0,0,0.1);"/>
</div>
<div class="column is-6">
<h3 class="title is-4">From Coordinate Matching to Intent Alignment</h3>
<p class="content">
We refine evaluation by checking if an action falls within the target element's <strong>bounding box</strong> rather than exact coordinate matching, better reflecting user intent and providing more reliable performance assessment.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="results">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Performance Results</h2>
<!-- Table 1 -->
<div class="columns is-centered" style="margin-top: 2rem;">
<div class="column is-four-fifths">
<figure class="image">
<img src="static/images/table1.jpg" alt="Table 1: Performance comparison of GUI agents on AndroidControl-Pro" style="border-radius: 8px; box-shadow: 0 4px 12px rgba(0,0,0,0.1);">
</figure>
<p class="has-text-centered content is-small" style="margin-top: 1em;">
<b>Table 1.</b> Performance comparison of GUI agents on AndroidControl-Pro. Grounding Accuracy (GA) for all models is evaluated using our proposed E<sub>bbox</sub>. The best results are in <b>bold</b>, and the second best are <u>underlined</u>.
</p>
</div>
</div>
<!-- Table 2 -->
<div class="columns is-centered" style="margin-top: 3rem;">
<div class="column is-four-fifths">
<figure class="image">
<img src="static/images/table2.jpg" alt="Table 2: Ablation analysis of the benchmark purification process" style="border-radius: 8px; box-shadow: 0 4px 12px rgba(0,0,0,0.1);">
</figure>
<p class="has-text-centered content is-small" style="margin-top: 1em;">
<b>Table 2.</b> Ablation analysis of the benchmark purification process on the Hard subset. SR Impr. (G) shows the SR gain from Origin-Bench to AndroidControl-Pro-G. SR Impr. (T) shows the SR gain from AndroidControl-Pro-G to the final AndroidControl-Pro. Best results are in <b>bold</b>, second best are <u>underlined</u>.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div class="bibtex-header">
<h2 class="title">BibTeX</h2>
<button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX to clipboard">
<i class="fas fa-copy"></i>
<span id="copy-text">Copy</span>
</button>
</div>
<pre id="bibtex-code"><code>@inproceedings{anonymous2024androidcontrolpro,
title={AndroidControl-Pro: Revealing the True Potential of GUI Agents through Benchmark Purification},
author={Anonymous},
booktitle={Under Review},
year={2024}
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content has-text-centered">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a>.
Website source code is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<script>
// Scroll to top functionality
function scrollToTop() {
window.scrollTo({top: 0, behavior: 'smooth'});
}
// Show/hide scroll button
window.onscroll = function() {
const scrollBtn = document.querySelector('.scroll-to-top');
if (document.body.scrollTop > 20 || document.documentElement.scrollTop > 20) {
scrollBtn.classList.add('visible');
} else {
scrollBtn.classList.remove('visible');
}
};
// Enhanced BibTeX copy function
function copyBibTeX() {
const bibtexCode = document.getElementById('bibtex-code');
const range = document.createRange();
range.selectNode(bibtexCode);
window.getSelection().removeAllRanges();
window.getSelection().addRange(range);
document.execCommand('copy');
window.getSelection().removeAllRanges();
const copyBtn = document.querySelector('.copy-bibtex-btn');
const copyText = document.getElementById('copy-text');
copyBtn.classList.add('copied');
copyText.textContent = 'Copied!';
setTimeout(() => {
copyBtn.classList.remove('copied');
copyText.textContent = 'Copy';
}, 2000);
}
</script>
</body>
</html>