-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
841 lines (766 loc) · 36.3 KB
/
index.html
File metadata and controls
841 lines (766 loc) · 36.3 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
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="description" content="ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents.">
<meta name="keywords" content="ToolCUA, Computer Use Agent, CUA, GUI Agent, Tool Use, OSWorld-MCP, GUI-Tool Path Orchestration">
<title>ToolCUA: Towards Optimal GUI-Tool Path Orchestration</title>
<link rel="icon" href="./assets/toolcua-logo.svg" type="image/svg+xml">
<link rel="stylesheet" href="./assets/bulma.min.css">
<style>
body {
font-family: "Inter", "Helvetica Neue", Helvetica, Arial, sans-serif;
color: #222;
}
.publication-title {
font-family: "Inter", "Helvetica Neue", Helvetica, Arial, sans-serif;
line-height: 1.15;
}
.publication-authors {
line-height: 1.45;
}
.publication-links {
margin-top: 1.25rem;
}
.link-block {
display: inline-block;
margin: 0.18rem 0.12rem;
}
.hero.teaser .hero-body {
padding-top: 1.5rem;
padding-bottom: 1.5rem;
}
.teaser img,
.paper-figure img {
width: 100%;
border-radius: 8px;
}
.paper-figure {
margin: 1.25rem 0 1.5rem;
}
.figure-caption {
color: #666;
font-size: 0.95rem;
margin-top: 0.7rem;
text-align: center;
}
.section {
padding: 2.4rem 1.5rem;
}
.section.compact {
padding-top: 1.5rem;
padding-bottom: 1.5rem;
}
.metric-box {
border: 1px solid #e5e5e5;
border-radius: 8px;
padding: 1.2rem 0.8rem;
height: 100%;
background: #fff;
}
.metric-value {
color: #6b46c1;
font-size: 2rem;
font-weight: 800;
line-height: 1.1;
}
.metric-label {
color: #666;
font-size: 0.95rem;
font-weight: 600;
margin-top: 0.35rem;
}
.comparison-table {
width: 100%;
border-collapse: separate;
border-spacing: 0;
font-size: 0.92rem;
color: #253044;
background: white;
border: 1px solid #e7e7e7;
border-radius: 8px;
overflow: hidden;
margin: 1.1rem 0 1.3rem;
}
.comparison-table th {
background: #ededed;
color: #111;
font-weight: 800;
padding: 0.7rem 0.75rem;
text-align: center;
border-bottom: 1px solid #d9d9d9;
}
.comparison-table td {
padding: 0.62rem 0.75rem;
text-align: center;
border-bottom: 1px solid #eeeeee;
}
.comparison-table td:first-child,
.comparison-table th:first-child {
text-align: left;
}
.comparison-table tr:last-child td {
border-bottom: none;
}
.comparison-table .our-model td {
background: #f0ebff;
font-weight: 800;
}
.placeholder-box {
border: 1px dashed #b8b8c8;
border-radius: 8px;
padding: 2rem;
color: #666;
background: #fafafa;
}
.case-showcase {
margin-top: 1.5rem;
}
.case-wide-container {
max-width: 1400px;
}
.case-container {
display: grid;
grid-template-columns: minmax(320px, 4fr) minmax(0, 6fr);
gap: 1rem;
align-items: stretch;
text-align: left;
}
.case-sidebar,
.case-panel {
border: 1px solid #e5e5e5;
border-radius: 8px;
background: #fff;
}
.case-sidebar {
overflow: hidden;
}
.case-instruction {
margin: 0 auto 1rem;
padding: 1rem 1.2rem;
border: 1px solid #e5e5e5;
border-radius: 8px;
background: #fafafa;
color: #333;
font-size: 1rem;
text-align: left;
}
.case-instruction strong {
margin-right: 0.45rem;
color: #111;
}
.case-step-list {
display: flex;
flex-wrap: wrap;
gap: 0.45rem;
padding: 0.8rem 1rem;
border-bottom: 1px solid #eeeeee;
}
.case-step {
border: 1px solid #dddddd;
border-radius: 999px;
background: #fff;
padding: 0.38rem 0.7rem;
text-align: center;
cursor: pointer;
flex: 0 0 auto;
}
.case-step:hover,
.case-step.is-active {
background: #f0ebff;
}
.case-step-number {
color: #6b46c1;
font-size: 0.82rem;
font-weight: 800;
text-transform: uppercase;
}
.case-step-title {
display: none;
color: #222;
font-size: 0.92rem;
font-weight: 650;
line-height: 1.35;
}
.case-main {
display: grid;
gap: 1rem;
min-width: 0;
}
.case-panel {
padding: 1rem;
min-width: 0;
}
.case-sidebar .case-panel {
border: 0;
border-radius: 0;
border-top: 1px solid #eeeeee;
}
.case-panel h4 {
color: #111;
font-size: 1rem;
font-weight: 800;
margin-bottom: 0.65rem;
}
.case-panel pre {
margin: 0;
max-height: 170px;
font-size: 0.8rem;
line-height: 1.45;
}
.case-current-step {
padding: 0.95rem 1rem 0;
color: #222;
font-size: 0.98rem;
font-weight: 750;
line-height: 1.35;
}
.case-state-panel {
padding: 1rem;
display: flex;
flex-direction: column;
min-height: 100%;
}
.case-state-frame {
width: 100%;
aspect-ratio: 16 / 9;
max-height: 760px;
border: 1px solid #eeeeee;
border-radius: 8px;
background: #111;
overflow: hidden;
display: flex;
align-items: center;
justify-content: center;
}
.case-state-image {
width: auto;
height: auto;
max-width: 100%;
max-height: 100%;
object-fit: contain;
display: block;
}
pre {
background: #f5f5f5;
border-radius: 8px;
color: #222;
overflow-x: auto;
padding: 1.2rem;
white-space: pre;
}
code {
color: inherit;
background: transparent;
}
.footer {
padding: 2rem 1.5rem;
}
@media screen and (max-width: 768px) {
.publication-title {
font-size: 2rem !important;
}
.comparison-table {
font-size: 0.78rem;
}
.comparison-table th,
.comparison-table td {
padding: 0.48rem 0.35rem;
}
.case-container {
grid-template-columns: 1fr;
}
.case-state-frame {
height: auto;
max-height: none;
}
}
</style>
</head>
<body>
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<div class="navbar-item has-dropdown is-hoverable">
<a class="navbar-link">More Research</a>
<div class="navbar-dropdown">
<a class="navbar-item" href="https://github.com/X-PLUG/MobileAgent">Mobile-Agent Series</a>
<a class="navbar-item" href="https://osworld-mcp.github.io/">OSWorld-MCP</a>
</div>
</div>
</div>
</div>
</nav>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">
<img src="assets/tongyi.png" width="40px" style="vertical-align: middle; margin-bottom: 10px;">
ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">Xuhao Hu<sup>2,1,*</sup>,</span>
<span class="author-block">Xi Zhang<sup>1,*</sup>,</span>
<span class="author-block">Haiyang Xu<sup>1,†</sup>,</span>
<span class="author-block">Kyle Qiao<sup>1</sup>,</span>
<span class="author-block">Jingyi Yang<sup>2</sup>,</span><br>
<span class="author-block">Xuanjing Huang<sup>2</sup>,</span>
<span class="author-block">Jing Shao<sup>3</sup>,</span>
<span class="author-block">Ming Yan<sup>1,†</sup>,</span>
<span class="author-block">Jieping Ye<sup>1</sup></span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Tongyi Lab, Alibaba Group</span>
<span class="author-block"><sup>2</sup>Fudan University</span>
<span class="author-block"><sup>3</sup>Shanghai Artificial Intelligence Laboratory</span>
</div>
<div class="is-size-6 publication-authors">
<span class="author-block"><sup>*</sup>Equal contribution</span>
<span class="author-block"><sup>†</sup>Corresponding author</span>
</div>
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/abs/2605.12481" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<svg aria-hidden="true" focusable="false" viewBox="0 0 24 24" width="18" height="18">
<path fill="currentColor" d="M6 2h8l5 5v15H6V2zm7 1.8V8h4.2L13 3.8zM8 11h8v1.6H8V11zm0 3.4h8V16H8v-1.6zm0 3.4h5v1.6H8v-1.6z"/>
</svg>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/X-PLUG/ToolCUA" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<svg aria-hidden="true" focusable="false" viewBox="0 0 24 24" width="18" height="18">
<path fill="currentColor" d="M12 2C6.48 2 2 6.58 2 12.22c0 4.52 2.87 8.35 6.84 9.71.5.09.68-.22.68-.49 0-.24-.01-.88-.01-1.73-2.78.62-3.37-1.37-3.37-1.37-.45-1.18-1.11-1.49-1.11-1.49-.91-.63.07-.62.07-.62 1 .07 1.53 1.05 1.53 1.05.9 1.56 2.35 1.11 2.92.85.09-.66.35-1.11.63-1.37-2.22-.26-4.56-1.14-4.56-5.06 0-1.12.39-2.03 1.03-2.75-.1-.26-.45-1.3.1-2.71 0 0 .84-.27 2.75 1.05A9.3 9.3 0 0 1 12 5.95c.85 0 1.7.12 2.5.34 1.91-1.32 2.75-1.05 2.75-1.05.55 1.41.2 2.45.1 2.71.64.72 1.03 1.63 1.03 2.75 0 3.93-2.34 4.79-4.57 5.05.36.32.68.94.68 1.9 0 1.37-.01 2.47-.01 2.81 0 .27.18.59.69.49A10.17 10.17 0 0 0 22 12.22C22 6.58 17.52 2 12 2z"/>
</svg>
</span>
<span>Code</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/mPLUG/ToolCUA-8B" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<img src="./assets/huggingface.svg" alt="Hugging Face" width="20" height="20">
</span>
<span>Model</span>
</a>
</span>
<span class="link-block">
<a href="#results" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<svg aria-hidden="true" focusable="false" viewBox="0 0 24 24" width="18" height="18">
<path fill="currentColor" d="M5 19h14v2H5v-2zm1-8h3v6H6v-6zm5-5h3v11h-3V6zm5 8h3v3h-3v-3z"/>
</svg>
</span>
<span>Results</span>
</a>
</span>
<span class="link-block">
<a href="#case-study" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<svg aria-hidden="true" focusable="false" viewBox="0 0 24 24" width="18" height="18">
<path fill="currentColor" d="M4 4h16v3H4V4zm0 5h16v3H4V9zm0 5h10v3H4v-3zm12 0h4v3h-4v-3z"/>
</svg>
</span>
<span>Cases</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<img src="./assets/main_teaser.png" alt="ToolCUA overview and main results teaser">
</div>
</div>
</section>
<section class="section compact" id="abstract">
<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>
Computer Use Agents (CUAs) are moving from text-only interaction toward real desktop operation, where agents must coordinate atomic GUI actions and high-level tool calls. However, simply exposing an agent to both action spaces does not make it reliable. In hybrid GUI-Tool environments, agents must learn when to continue through visual GUI grounding, when to invoke structured tools, and when tool calls may actually hurt task success.
</p>
<p>
We introduce <b>ToolCUA</b>, an end-to-end CUA designed for <b>optimal GUI-Tool path selection</b>. ToolCUA scales interleaved GUI-Tool trajectories from existing GUI-only data, applies Tool-Bootstrapped GUI RFT to acquire tool knowledge and calibrate local switching behavior, and further optimizes full task trajectories with Online Agentic RL using a Tool-Efficient Path Reward.
</p>
<p>
On OSWorld-MCP, ToolCUA-8B achieves <b>46.85%</b> accuracy, a relative improvement of about <b>66%</b> over Qwen3-VL-8B-Instruct, while reaching the lowest average completion steps among compared models at <b>14.93</b>.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="problem">
<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">Path Selection Confusion Under Hybrid Actions</h2>
<div class="content has-text-justified">
<p>
Giving CUAs both GUI actions and tool calls does not automatically make them better. Our diagnostic study exposes a direct <b>path selection confusion</b> problem under hybrid actions: some models stay GUI-centric and almost never call tools, while stronger models may overuse tools, shorten trajectories, and still lose task success.
</p>
</div>
<div class="paper-figure">
<img src="./assets/gui_tool_confusion.png" alt="GUI-Tool path confusion">
<div class="figure-caption">Current CUAs suffer from path selection confusion when GUI actions and tools are available together.</div>
</div>
<div class="content has-text-justified">
<p>
Qwen3-VL-8B barely invokes tools after they are introduced, with only 0.003 tool calls per trajectory and an accuracy drop from 29.0% to 28.2%. Qwen3-VL-235B calls tools much more frequently, reducing average steps from 25.9 to 17.4, but accuracy still drops from 41.1% to 38.1%. The core bottleneck is therefore whether the agent can choose the right GUI-Tool path at each state.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="method">
<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">ToolCUA Method</h2>
<div class="paper-figure">
<img src="./assets/method_overview.png" alt="ToolCUA method overview">
<div class="figure-caption">ToolCUA scales interleaved GUI-Tool trajectories, calibrates switching decisions, and optimizes trajectory-level paths with online RL.</div>
</div>
<div class="content has-text-justified">
<p>
<b>Interleaved GUI-Tool Trajectory Scaling.</b> ToolCUA starts from existing GUI-only trajectories and uses MLLMs to synthesize trajectory-aware grounded tools. It then generates tool trajectories with next-state grounding and constructs diverse interleaved GUI-Tool trajectories by replacing suitable GUI subsequences with tool calls.
</p>
<p>
<b>Tool-Bootstrapped GUI RFT.</b> Warmup SFT teaches tool schemas, arguments, responses, and GUI state transitions. Single-turn RL on critical switching steps further calibrates local GUI-versus-Tool decisions, so the model learns when a tool is appropriate in context.
</p>
<p>
<b>Online Agentic RL.</b> ToolCUA then performs long-horizon rollout in GUI-Tool environments. Its Tool-Efficient Path Reward combines task success, format validity, tool appropriateness, and path efficiency, encouraging agents to use tools when helpful, avoid them when unnecessary, and complete successful tasks in fewer steps.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="data">
<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">Scaling GUI-Tool Data from GUI-Only Trajectories</h2>
<div class="content has-text-justified">
<p>
High-quality GUI-Tool trajectories are scarce because real tools are application-specific, costly to build, and hard to maintain. ToolCUA repurposes GUI-only demonstrations into hybrid supervision, enabling the model to learn not only how to call tools, but also where tools should replace, complement, or defer to GUI operations.
</p>
</div>
<div class="paper-figure">
<img src="./assets/synthesis_example.png" alt="Synthetic interleaved GUI-Tool trajectory example">
<div class="figure-caption">Example of converting GUI-only operation traces into interleaved GUI-Tool supervision.</div>
</div>
<div class="columns is-multiline">
<div class="column is-one-third"><div class="metric-box"><div class="metric-value">10k</div><div class="metric-label">source trajectories</div></div></div>
<div class="column is-one-third"><div class="metric-box"><div class="metric-value">192k</div><div class="metric-label">raw GUI steps</div></div></div>
<div class="column is-one-third"><div class="metric-box"><div class="metric-value">180k</div><div class="metric-label">SFT steps</div></div></div>
<div class="column is-one-third"><div class="metric-box"><div class="metric-value">5k</div><div class="metric-label">critical switching steps</div></div></div>
<div class="column is-one-third"><div class="metric-box"><div class="metric-value">4,350</div><div class="metric-label">unique tools</div></div></div>
<div class="column is-one-third"><div class="metric-box"><div class="metric-value">19.75</div><div class="metric-label">avg. tools per trajectory</div></div></div>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="results">
<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">Main Results on OSWorld-MCP</h2>
<div class="content has-text-justified">
<p>
ToolCUA is evaluated on OSWorld-MCP, which extends OSWorld with GUI actions and 150+ MCP tools across realistic desktop applications. We report Accuracy, Tool Invocation Rate (TIR), and Average Completion Steps (ACS) over 333 feasible tasks.
</p>
</div>
<div class="table-container">
<table class="comparison-table">
<thead>
<tr>
<th>Agent Model</th>
<th>Accuracy ↑</th>
<th>TIR ↑</th>
<th>ACS ↓</th>
</tr>
</thead>
<tbody>
<tr><td>Claude-4-Sonnet</td><td>43.54</td><td>35.74</td><td>19.76</td></tr>
<tr><td>Gemini-3.1-Pro</td><td>41.14</td><td>34.23</td><td>25.40</td></tr>
<tr><td>Claude-4.5-Sonnet</td><td>48.35</td><td>40.24</td><td>19.07</td></tr>
<tr><td>GUI-Owl-1.5-8B</td><td>43.84</td><td>36.04</td><td>21.19</td></tr>
<tr><td>GUI-Owl-1.5-32B</td><td>48.05</td><td>41.14</td><td>24.19</td></tr>
<tr><td>Qwen3-VL-8B-Instruct</td><td>28.23</td><td>8.41</td><td>19.34</td></tr>
<tr class="our-model"><td>ToolCUA-8B</td><td>46.85</td><td>24.32</td><td>14.93</td></tr>
</tbody>
</table>
</div>
<div class="paper-figure">
<img src="./assets/main_results.png" alt="ToolCUA main OSWorld-MCP results">
<div class="figure-caption">ToolCUA-8B achieves strong accuracy while producing the shortest average successful execution paths.</div>
</div>
<div class="content has-text-justified">
<p>
Compared with Qwen3-VL-8B-Instruct, ToolCUA improves overall accuracy by +18.62 points, raises TIR from 8.41 to 24.32, and reduces ACS from 19.34 to 14.93. This shows that ToolCUA learns selective tool usage rather than simply increasing tool calls.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="analysis">
<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">Path Selection, Generalization, and Ablations</h2>
<div class="content has-text-justified">
<p>
ToolCUA demonstrates strong generalization beyond the training distribution. Although online agentic RL is conducted only on single-application Linux tasks and excludes the <i>multi_apps</i> category, ToolCUA improves on the held-out <i>multi_apps</i> domain from the baseline 9.8% and the pre-online-RL stage 18.5% to 23.9%. It also achieves consistent gains across specialized domains, including <i>libreoffice_calculation</i> and <i>vs_code</i>.
</p>
<p>
Beyond cross-task transfer, ToolCUA further generalizes to unseen Windows desktop environments. Despite being trained on Linux-based trajectories and sandboxes, ToolCUA reaches 33.8% accuracy on WindowsAgentArena, outperforming the Qwen3-VL-8B-Instruct baseline by 7.4 percentage points and surpassing larger Qwen3-VL variants.
</p>
</div>
<div class="paper-figure">
<img src="./assets/app_results.png" alt="ToolCUA app and OOD results">
<div class="figure-caption">ToolCUA generalizes across multi-app workflows, application domains, and unseen Windows desktop tasks.</div>
</div>
<div class="content has-text-justified">
<p>
<b>The Importance of Interleaved GUI-Tool Trajectory Data.</b> Removing coldstart RFT with synthetic interleaved GUI-Tool data and directly applying online agentic RL still improves task accuracy, but the model struggles to acquire reliable tool-calling behavior. Its TIR remains consistently low and tool calls stay close to zero throughout most of training, showing that online RL with tool-efficiency rewards alone is insufficient to overcome the GUI-centric bias of base models.
</p>
<p>
<b>Advantages of Tool-Efficient Path Reward.</b> Replacing our path reward with vanilla multi-turn GRPO makes the accuracy curve less stable, produces a clear drop around steps 8--11, and leaves an eventual gap of about 7 percentage points after 20 training steps. TIR and tool calls fluctuate without a consistent upward trend, while trajectory length lacks a stable downward trend, validating that the Tool-Efficient Path Reward is essential for tool-appropriate and efficiency-aware path selection.
</p>
</div>
<div class="paper-figure">
<img src="./assets/rl_dynamics.png" alt="ToolCUA RL dynamics and ablations">
<div class="figure-caption">Online Agentic RL dynamics show that both hybrid data and path-aware rewards are necessary for stable GUI-Tool orchestration.</div>
</div>
<div class="content has-text-justified">
<p>
<b>Hybrid GUI-Tool Training is More Effective than Pure GUI.</b> A pure GUI pipeline improves the baseline from 29.03% to 34.93% after SFT and 42.05% after agentic RL, but both stages remain below their GUI-Tool counterparts. RFT training with synthetic interleaved GUI-Tool trajectories already reaches 38.13%, and full ToolCUA further improves to 46.85%, indicating that hybrid action spaces provide a more effective training environment for learning when structured tool calls can replace redundant low-level GUI operations.
</p>
</div>
<div class="paper-figure">
<img src="./assets/more_results.png" alt="ToolCUA pure GUI comparison and WindowsAgentArena generalization">
<div class="figure-caption">Hybrid GUI-Tool training outperforms pure GUI training, and ToolCUA transfers to Windows desktop environments.</div>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="case-study">
<div class="container case-wide-container">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-3">Case Study</h2>
<p>
Click a case and a trajectory step to inspect the agent action, tool response, and post-execution desktop state.
</p>
<div class="case-showcase">
<div class="tabs is-centered is-toggle">
<ul>
<li class="case-tab is-active" data-case-id="calc"><a>LibreOffice Calc</a></li>
<li class="case-tab" data-case-id="vscode"><a>VS Code</a></li>
</ul>
</div>
<div class="case-instruction">
<strong>Task Instruction</strong>
<span id="case-instruction"></span>
</div>
<div class="case-container">
<div class="case-sidebar">
<div id="case-step-list" class="case-step-list"></div>
<div id="case-current-step" class="case-current-step"></div>
<div class="case-panel">
<h4>Agent Action</h4>
<pre><code id="case-action"></code></pre>
</div>
<div class="case-panel">
<h4>Tool Response</h4>
<pre><code id="case-response"></code></pre>
</div>
</div>
<div class="case-main">
<div class="case-panel case-state-panel">
<h4>Post-Execution State</h4>
<div class="case-state-frame">
<img id="case-state" class="case-state-image" src="./assets/cases/ex2_step_2.png" alt="Case post-execution state">
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{hu2026toolcua,
title={ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents},
author={Hu, Xuhao and Zhang, Xi and Xu, Haiyang and Qiao, Kyle and Yang, Jingyi and Huang, Xuanjing and Shao, Jing and Yan, Ming and Ye, Jieping},
journal={arXiv preprint arXiv:2605.12481},
year={2026}
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International
License</a>.
</p>
<p>
This means you are free to borrow the <a href="https://github.com/nerfies/nerfies.github.io">source
code</a> of this website,
we just ask that you link back to this page in the footer.
Please remember to remove the analytics code included in the
header of the website which
you do not want on your website.
</p>
</div>
</div>
</div>
</div>
</footer>
<script>
const caseData = {
calc: {
instruction: 'Create two pivot tables in a new sheet named Sheet2 showing the total revenue for each product and sales channel.',
steps: [
{
step: 'Step 2',
title: 'Get workbook information to identify the file path and sheet details.',
action: `Action: Get the workbook information to identify the file path and sheet details.\n\n<tool_call>\n{"name": "osworld_mcp_libreoffice_calc.get_workbook_info", "arguments": {}}\n</tool_call>`,
response: `<tool_response>\n{"type": "text", "text": "{\\"success\\": true, \\"result\\": {\\"file_path\\": \\"file:///home/user/SummerSales.xlsx\\", \\"file_title\\": \\"SummerSales.xlsx\\", \\"sheets\\": [{\\"name\\": \\"Sheet1\\", \\"index\\": 0, \\"visible\\": true, \\"row_count\\": 30, \\"column_count\\": 7}], \\"active_sheet\\": \\"Sheet1\\", \\"sheet_count\\": 1}, \\"error_message\\": null}"}\n</tool_response>`,
image: './assets/cases/ex2_step_2.png'
},
{
step: 'Step 3',
title: 'Get content information of Sheet1 to understand headers and data ranges.',
action: `Action: Get content information of Sheet1 to understand the data structure including headers and data ranges.\n\n<tool_call>\n{"name": "osworld_mcp_libreoffice_calc.env_info", "arguments": {"sheet_name": "Sheet1"}}\n</tool_call>`,
response: `<tool_response>\n{"type": "text", "text": "{\\"success\\": true, \\"result\\": {\\"name\\": \\"Sheet1\\", \\"data\\": \\" | A| B| C | D | E | F | G |\\\\n--|-----...}, \\"error_message\\": null}"}\n</tool_response>`,
image: './assets/cases/ex2_step_3.png'
},
{
step: 'Step 4',
title: 'Create a new sheet named Sheet2 for the pivot tables.',
action: `Action: Create a new sheet named "Sheet2" to serve as the destination for the pivot tables.\n\n<tool_call>\n{"name": "osworld_mcp_libreoffice_calc.switch_active_sheet", "arguments": {"sheet_name": "Sheet2"}}\n</tool_call>`,
response: `<tool_response>\n{"type": "text", "text": "{\\"success\\": true, \\"result\\": \\"Success\\", \\"error_message\\": null}"}\n</tool_response>`,
image: './assets/cases/ex2_step_4.png'
},
{
step: 'Step 5',
title: 'Create PivotTable1 for total revenue by product.',
action: `Action: Create the first pivot table named 'PivotTable1' in Sheet2 to show total revenue for each product.\n\n<tool_call>\n{"name": "osworld_mcp_libreoffice_calc.create_pivot_table", "arguments": {"source_sheet": "Sheet1", "table_name": "PivotTable1", "row_fields": ["F"], "value_fields": ["G"], "aggregation_function": "sum", "target_cell": "A1"}}\n</tool_call>`,
response: `<tool_response>\n{"success": true, "result": "Success", "error_message": null}\n</tool_response>`,
image: './assets/cases/ex2_step_5.png'
},
{
step: 'Step 6',
title: 'Create PivotTable2 for total revenue by sales channel.',
action: `Action: Create the second pivot table named 'PivotTable2' in Sheet2 to show total revenue for each sales channel.\n\n<tool_call>\n{"name": "osworld_mcp_libreoffice_calc.create_pivot_table", "arguments": {"source_sheet": "Sheet1", "table_name": "PivotTable2", "row_fields": ["E"], "value_fields": ["G"], "aggregation_function": "sum", "target_cell": "A19"}}\n</tool_call>`,
response: `<tool_response>\n{"success": true, "result": "Success", "error_message": null}\n</tool_response>`,
image: './assets/cases/ex2_step_6.png'
}
]
},
vscode: {
instruction: 'Add folder "/home/user/data1" and folder "/home/user/data2" to the current VS Code workspace.',
steps: [
{
step: 'Step 1',
title: 'Add /home/user/data1 to the current workspace.',
action: `Action: Add the folder '/home/user/data1' to the current workspace in VSCode.\n\n<tool_call>\n{"name": "osworld_mcp_code.add_folder", "arguments": {"folder": "/home/user/data1"}}\n</tool_call>`,
response: `<tool_response>\n{"type": "text", "text": "{\\"success\\": true, \\"result\\": \\"Successfully added folder\\", \\"error_message\\": null}"}\n</tool_response>`,
image: './assets/cases/ex4_step_1.png'
},
{
step: 'Step 2',
title: 'Add /home/user/data2 to the current workspace.',
action: `Action: Add the folder '/home/user/data2' to the current workspace in VSCode.\n\n<tool_call>\n{"name": "osworld_mcp_code.add_folder", "arguments": {"folder": "/home/user/data2"}}\n</tool_call>`,
response: `<tool_response>\n{"type": "text", "text": "{\\"success\\": true, \\"result\\": \\"Successfully added folder\\", \\"error_message\\": null}"}\n</tool_response>`,
image: './assets/cases/ex4_step_2.png'
},
{
step: 'Step 3',
title: 'Click the trust dialog to finalize workspace access.',
action: `Action: Click on the "Yes, I trust the authors" button in the trust dialog popup.\n\n<tool_call>\n{"name": "computer_use", "arguments": {"action": "click", "coordinate": [596, 627]}}\n</tool_call>`,
response: '## Not available',
image: './assets/cases/ex4_step_3.png'
},
{
step: 'Step 4',
title: 'Terminate the task as successful.',
action: `Action: Terminate the task as successful since the goal has been accomplished.\n\n<tool_call>\n{"name": "computer_use", "arguments": {"action": "terminate", "status": "success"}}\n</tool_call>`,
response: '## Not available',
image: './assets/cases/ex4_step_4.png'
}
]
}
};
let activeCase = 'calc';
let activeStep = 0;
function renderCase() {
const currentCase = caseData[activeCase];
const stepList = document.getElementById('case-step-list');
document.getElementById('case-instruction').textContent = currentCase.instruction;
stepList.innerHTML = '';
currentCase.steps.forEach((item, index) => {
const button = document.createElement('button');
button.className = `case-step${index === activeStep ? ' is-active' : ''}`;
button.type = 'button';
button.innerHTML = `<span class="case-step-number">${item.step}</span><span class="case-step-title">${item.title}</span>`;
button.addEventListener('click', () => {
activeStep = index;
renderCase();
});
stepList.appendChild(button);
});
const currentStep = currentCase.steps[activeStep];
document.getElementById('case-current-step').textContent = `${currentStep.step}: ${currentStep.title}`;
document.getElementById('case-action').textContent = `${currentStep.step}: ${currentStep.action}`;
document.getElementById('case-response').textContent = currentStep.response;
const stateImage = document.getElementById('case-state');
stateImage.src = currentStep.image;
stateImage.alt = `${activeCase} ${currentStep.step} post-execution state`;
}
document.querySelectorAll('.case-tab').forEach((tab) => {
tab.addEventListener('click', () => {
document.querySelectorAll('.case-tab').forEach((item) => item.classList.remove('is-active'));
tab.classList.add('is-active');
activeCase = tab.dataset.caseId;
activeStep = 0;
renderCase();
});
});
renderCase();
</script>
</body>
</html>