-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
2784 lines (2496 loc) · 125 KB
/
index.html
File metadata and controls
2784 lines (2496 loc) · 125 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
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=5, user-scalable=yes">
<!-- Google Search Console Verification -->
<meta name="google-site-verification" content="5XSefLo9dX0I_Szro49mP5w54fCMDDuxB3E3LViw5mU" />
<!-- Canonical URL -->
<link rel="canonical" href="https://visgym.github.io/" />
<!-- Primary Meta Tags -->
<meta name="title" content="VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents - Zirui Wang, Junyi Zhang, Jiaxin Ge, Long Lian, Letian Fu, Lisa Dunlap, Ken Goldberg, Xudong Wang, Ion Stoica, David M. Chan, Sewon Min, Joseph E. Gonzalez">
<meta name="description" content="VisGym consists of 17 diverse, long-horizon environments designed to systematically evaluate, diagnose, and train VLMs on visually interactive tasks with different domains, levels of state observability, and types of observations.">
<meta name="keywords" content="VisGym,Vis Gym, Visual Gym, Vision Gym, Vis Gym Benchmark, Visual-Agent Gym, Multimodal Gym, VLM Gym, Interactive Visual Environments, Visual Reasoning Environments, Long-Horizon Tasks, Visual Interactive Tasks, Multimodal Benchmark, VLM Benchmark, Gym Environment Benchmark, Vision-Language Agents, Multimodal Agents, Vision Language Models, Interactive RL Environments, CV Benchmark Environments, Visual Interaction Benchmark">
<meta name="author" content="Zirui Wang, Junyi Zhang, Jiaxin Ge, Long Lian, Letian Fu, Lisa Dunlap, Ken Goldberg, Xudong Wang, Ion Stoica, David M. Chan, Sewon Min, Joseph E. Gonzalez">
<meta name="robots" content="index, follow">
<meta name="language" content="English">
<!-- Open Graph / Facebook -->
<meta property="og:type" content="article">
<meta property="og:site_name" content="VisGym Research">
<meta property="og:title" content="VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents">
<meta property="og:description" content="VisGym consists of 17 diverse, long-horizon environments designed to systematically evaluate, diagnose, and train VLMs on visually interactive tasks with different domains, levels of state observability, and types of observations.">
<meta property="og:url" content="https://visgym.github.io/">
<meta property="og:image" content="https://visgym.github.io/static/images/teaser.png">
<meta property="og:image:width" content="1200">
<meta property="og:image:height" content="630">
<meta property="og:image:alt" content="VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents - Research Preview">
<meta property="article:published_time" content="2024-12-19T00:00:00.000Z">
<meta property="article:author" content="Zirui Wang">
<meta property="article:section" content="Research">
<meta property="article:tag" content="Vision Language Models">
<meta property="article:tag" content="Multimodal Agents">
<!-- Twitter -->
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:site" content="@bair_lab">
<meta name="twitter:creator" content="@bair_lab">
<meta name="twitter:title" content="VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents">
<meta name="twitter:description" content="VisGym consists of 17 diverse, long-horizon environments designed to systematically evaluate, diagnose, and train VLMs on visually interactive tasks.">
<meta name="twitter:image" content="https://visgym.github.io/static/images/teaser.png">
<meta name="twitter:image:alt" content="VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents - Research Preview">
<!-- Academic/Research Specific -->
<meta name="citation_title" content="VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents">
<meta name="citation_author" content="Wang, Zirui">
<meta name="citation_author" content="Zhang, Junyi">
<meta name="citation_author" content="Ge, Jiaxin">
<meta name="citation_author" content="Lian, Long">
<meta name="citation_author" content="Fu, Letian">
<meta name="citation_author" content="Dunlap, Lisa">
<meta name="citation_author" content="Goldberg, Ken">
<meta name="citation_author" content="Wang, Xudong">
<meta name="citation_author" content="Stoica, Ion">
<meta name="citation_author" content="Chan, David M.">
<meta name="citation_author" content="Min, Sewon">
<meta name="citation_author" content="Gonzalez, Joseph E.">
<meta name="citation_publication_date" content="2024">
<meta name="citation_conference_title" content="Under Review">
<meta name="citation_pdf_url" content="https://arxiv.org/pdf/2601.16973">
<!-- Additional SEO -->
<meta name="theme-color" content="#2563eb">
<meta name="msapplication-TileColor" content="#2563eb">
<!-- 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://ajax.googleapis.com">
<link rel="preconnect" href="https://documentcloud.adobe.com">
<link rel="preconnect" href="https://cdn.jsdelivr.net">
<title>VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents | VisGym Research</title>
<!-- Favicon -->
<link rel="icon" type="image/x-icon" href="static/images/favicon.ico">
<!-- Critical CSS - Load synchronously -->
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/index.css">
<!-- Non-critical CSS - Load asynchronously -->
<link rel="preload" href="static/css/bulma-carousel.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="static/css/bulma-slider.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="static/css/fontawesome.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="static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="static/css/bulma-slider.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
</noscript>
<!-- Fonts - Optimized loading -->
<link href="https://fonts.googleapis.com/css2?family=Karla:wght@400;500;600;700&display=swap" rel="stylesheet">
<link href="https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@400;500;600;700;800&display=swap" rel="stylesheet">
<link href="https://fonts.googleapis.com/css2?family=Space+Mono:wght@400;700&display=swap" rel="stylesheet">
<!-- ECharts for Radar Chart -->
<script src="https://cdn.jsdelivr.net/npm/echarts/dist/echarts.min.js"></script>
<!-- Tailwind CSS for key_insights and failure_analysis -->
<script src="https://cdn.tailwindcss.com"></script>
<!-- Styles for embedded sections -->
<style>
/* Task Success Rate Styles */
.task-success-rate-section {
background-color: transparent;
}
.task-success-rate-section body {
font-family: 'Karla', -apple-system, BlinkMacSystemFont, sans-serif;
background-color: transparent;
padding: 0;
margin: 0;
}
.task-success-rate-section .table-container {
background: transparent;
border-radius: 0;
box-shadow: none;
padding: 0.5rem 0 0 0;
margin: 0 0 0 0;
max-width: 100%;
width: 100%;
}
.task-success-rate-section .table-header {
margin-bottom: 0;
position: relative;
overflow: visible;
}
.task-success-rate-section .section-main-title {
font-family: 'Space Grotesk', sans-serif;
font-size: 2rem;
font-weight: 700;
color: #003262;
margin-bottom: 0;
}
.task-success-rate-section .table-controls {
display: flex;
gap: 1rem;
margin-bottom: 1.5rem;
flex-wrap: wrap;
align-items: center;
position: relative;
overflow: visible;
}
.task-success-rate-section .task-selector {
width: 100%;
}
.task-success-rate-section .task-selector-header {
transition: background-color 0.3s;
}
.task-success-rate-section .task-selector-header:hover {
background-color: #f0f0f0;
}
.task-success-rate-section .task-selector-content {
animation: slideDown 0.3s ease-out;
overflow: visible;
}
@keyframes slideDown {
from {
opacity: 0;
transform: translateY(-8px);
}
to {
opacity: 1;
transform: translateY(0);
}
}
.task-success-rate-section .task-checkboxes-container {
display: flex;
flex-direction: column;
flex-wrap: wrap;
max-height: 300px;
gap: 0.5rem;
padding: 1rem;
border: 2px solid #e0e0e0;
border-radius: 6px;
background-color: white;
overflow-y: auto;
overflow-x: hidden;
}
.task-success-rate-section .task-checkbox-item {
display: flex;
align-items: center;
gap: 0.5rem;
padding: 0.3rem 0;
cursor: pointer;
user-select: none;
}
.task-success-rate-section .task-checkbox-item:hover {
background-color: #f5f5f5;
border-radius: 4px;
}
.task-success-rate-section .task-checkbox-item input[type="checkbox"] {
width: 18px;
height: 18px;
cursor: pointer;
accent-color: #003262;
}
.task-success-rate-section .task-checkbox-item label {
font-family: 'Karla', sans-serif;
font-size: 0.9rem;
color: #333;
cursor: pointer;
margin: 0;
flex: 1;
}
.task-success-rate-section .filter-btn {
font-family: 'Space Mono', monospace;
font-size: 0.85rem;
padding: 0.4rem 0.8rem;
border: 2px solid #e0e0e0;
background: white;
border-radius: 6px;
cursor: pointer;
transition: all 0.3s;
}
.task-success-rate-section .filter-btn:hover {
border-color: #003262;
background: #f5f5f5;
}
.task-success-rate-section .success-rate-table {
width: 100%;
border-collapse: collapse;
font-size: 0.85rem;
overflow-x: auto;
display: block;
min-width: 2000px;
}
.task-success-rate-section .success-rate-table thead {
background: #f8f9fa;
position: sticky;
top: 0;
z-index: 10;
}
.task-success-rate-section .success-rate-table th {
font-family: 'Space Grotesk', sans-serif;
font-weight: 600;
padding: 0.75rem 0.5rem;
text-align: left;
border-bottom: 2px solid #e0e0e0;
border-right: 1px dashed #e0e0e0;
color: #003262;
cursor: pointer;
user-select: none;
white-space: nowrap;
position: relative;
font-size: 0.7rem;
}
.task-success-rate-section .success-rate-table th.sortable::after {
content: ' ↕';
opacity: 0.3;
font-size: 0.8em;
}
.task-success-rate-section .success-rate-table th.sort-asc::after {
content: ' ↑';
opacity: 1;
color: #003262;
}
.task-success-rate-section .success-rate-table th.sort-desc::after {
content: ' ↓';
opacity: 1;
color: #003262;
}
.task-success-rate-section .success-rate-table td {
padding: 0.6rem 0.5rem;
border-bottom: 1px solid #f0f0f0;
border-right: 1px dashed #e0e0e0;
white-space: nowrap;
}
.task-success-rate-section .success-rate-table tbody tr:nth-child(odd) {
background-color: #f5f5f5;
}
.task-success-rate-section .success-rate-table tbody tr:nth-child(even) {
background-color: #ffffff;
}
.task-success-rate-section .success-rate-table tbody tr:hover {
background-color: #e8e8e8;
}
.task-success-rate-section .success-rate-table tbody tr.hidden {
display: none;
}
.task-success-rate-section .success-rate-table tbody tr.ours-model {
background-color: #FEF5E7 !important;
}
.task-success-rate-section .success-rate-table tbody tr.ours-model:hover {
background-color: #FDE8C8 !important;
}
.task-success-rate-section .success-rate-table tbody tr.ours-model td {
background-color: #FEF5E7 !important;
}
.task-success-rate-section .task-name {
font-weight: 600;
color: #333;
font-family: 'Space Mono', monospace;
font-size: 0.85rem;
position: sticky;
left: 0;
background-color: white;
z-index: 5;
box-shadow: 2px 0 4px rgba(0,0,0,0.1);
}
.task-success-rate-section .success-rate-table thead th:first-child {
position: sticky;
left: 0;
background-color: #f8f9fa;
z-index: 15;
box-shadow: 2px 0 4px rgba(0,0,0,0.1);
}
.task-success-rate-section .success-rate-table tbody td:first-child {
position: sticky;
left: 0;
background-color: white;
z-index: 5;
box-shadow: 2px 0 4px rgba(0,0,0,0.1);
}
.task-success-rate-section .success-rate-table tbody tr:nth-child(odd) td:first-child {
background-color: #f5f5f5;
}
.task-success-rate-section .success-rate-table tbody tr.ours-model td:first-child {
background-color: #FEF5E7 !important;
}
.task-success-rate-section .rate-value.high {
color: #3B7EA1;
}
.task-success-rate-section .rate-value.medium {
color: #F48024;
}
.task-success-rate-section .rate-value.low {
color: #ED4E33;
}
@media (max-width: 768px) {
.task-success-rate-section .section-main-title {
font-size: 1.5rem !important;
}
.task-success-rate-section .task-selector-content {
position: absolute !important;
top: 100% !important;
left: 0 !important;
width: 100% !important;
z-index: 9999 !important;
margin-top: 4px !important;
background-color: white !important;
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.15) !important;
}
}
/* Key Insights Styles */
.key-insights-section .berkeley-blue { color: #003262; }
.key-insights-section .berkeley-bg { background-color: #003262; }
/* Failure Analysis Styles */
.failure-analysis-section .berkeley-blue { color: #003262; }
.failure-analysis-section .bg-berkeley-blue { background-color: #003262; }
.failure-analysis-section .font-mono { font-family: 'Space Mono', monospace; }
.failure-analysis-section #stringsightImage,
.failure-analysis-section #stringsightImage2,
.failure-analysis-section #stringsightImage3,
.failure-analysis-section #stringsightImage4 {
image-rendering: -webkit-optimize-contrast;
image-rendering: crisp-edges;
max-width: 100%;
max-height: 100%;
transition: opacity 0.6s ease-in-out;
}
.failure-analysis-section #stringsightCarousel,
.failure-analysis-section #stringsightCarousel2,
.failure-analysis-section #stringsightCarousel3,
.failure-analysis-section #stringsightCarousel4 {
display: flex;
align-items: center;
justify-content: center;
width: 100%;
height: 100%;
min-height: 200px;
}
</style>
<!-- Defer non-critical JavaScript -->
<script defer src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script defer src="static/js/bulma-carousel.min.js"></script>
<script defer src="static/js/bulma-slider.min.js"></script>
<script defer src="static/js/index.js"></script>
<script defer src="static/js/radar-chart.js"></script>
<!-- Structured Data for Academic Papers -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "ScholarlyArticle",
"headline": "VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents",
"alternateName": [
"Vis Gym",
"Vision Gym",
"Visual Gym",
"VisGym Benchmark",
"Visual Gym Benchmark"
],
"description": "VisGym consists of 17 diverse, long-horizon environments designed to systematically evaluate, diagnose, and train VLMs on visually interactive tasks with different domains, levels of state observability, and types of observations.",
"author": [
{
"@type": "Person",
"name": "Zirui Wang",
"affiliation": {
"@type": "Organization",
"name": "UC Berkeley"
}
},
{
"@type": "Person",
"name": "Junyi Zhang",
"affiliation": {
"@type": "Organization",
"name": "UC Berkeley"
}
},
{
"@type": "Person",
"name": "Jiaxin Ge",
"affiliation": {
"@type": "Organization",
"name": "UC Berkeley"
}
}
],
"datePublished": "2025-11-19",
"publisher": {
"@type": "Organization",
"name": "Under Review"
},
"url": "https://visgym.github.io/",
"image": "https://visgym.github.io/static/images/teaser.png",
"keywords": ["Vision Language Models", "Multimodal Agents", "Reinforcement Learning", "machine learning", "computer vision", "Interactive Environments", "VLM Evaluation"],
"abstract": "Modern Vision–Language Models (VLMs) remain poorly characterized in multi-step visual interactions, particularly in how they integrate perception, memory, and action over long horizons. We introduce VisGym, a gymnasium of 17 environments for evaluating and training VLMs. The suite spans symbolic puzzles, real-image understanding, navigation, and manipulation, and provides flexible controls over difficulty, input representation, planning horizon, and feedback. We also provide multi-step solvers that generate structured demonstrations, enabling supervised finetuning. Our evaluations show that all frontier models struggle in interactive settings, achieving low success rates in both the easy (46.6%) and hard (26.0%) configurations. Our experiments reveal notable limitations: models struggle to effectively leverage long context, performing worse with an unbounded history than with truncated windows. Furthermore, we find that several text-based symbolic tasks become substantially harder once rendered visually. However, explicit goal observations, textual feedback, and exploratory demonstrations in partially observable or unknown-dynamics settings for supervised finetuning yield consistent gains, highlighting concrete failure modes and pathways for improving multi-step visual decision-making.",
"isAccessibleForFree": true,
"license": "https://creativecommons.org/licenses/by/4.0/",
"mainEntity": {
"@type": "WebPage",
"@id": "https://visgym.github.io/"
},
"about": [
{
"@type": "Thing",
"name": "Vision Language Models"
},
{
"@type": "Thing",
"name": "Multimodal Agents"
},
{
"@type": "Thing",
"name": "Interactive Environments"
},
{
"@type": "Thing",
"name": "Visual Gym Benchmark"
}
]
}
</script>
<!-- Website/Organization Structured Data -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "BAIR Lab, UC Berkeley",
"url": "https://bair.berkeley.edu/",
"logo": "https://visgym.github.io/static/images/favicon.ico",
"sameAs": [
"https://twitter.com/bair_lab",
"https://github.com/visgym"
]
}
</script>
<!-- Google Tag Manager -->
<script>(function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':
new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],
j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src=
'https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);
})(window,document,'script','dataLayer','GTM-NX9M2F8W');</script>
<!-- End Google Tag Manager -->
<!-- Cloudflare Web Analytics -->
<script defer src='https://static.cloudflareinsights.com/beacon.min.js' data-cf-beacon='{"token": "bec8fd563f2c49e28c16f20dca243c43"}'></script>
<!-- End Cloudflare Web Analytics -->
</head>
<body>
<!-- Google Tag Manager (noscript) -->
<noscript><iframe src="https://www.googletagmanager.com/ns.html?id=GTM-NX9M2F8W"
height="0" width="0" style="display:none;visibility:hidden"></iframe></noscript>
<!-- End Google Tag Manager (noscript) -->
<!-- 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">
<!-- Apple-style Split Layout Hero Section -->
<section class="hero apple-hero">
<!-- Full-width Title -->
<div class="apple-title-container page-shell">
<h1 class="publication-title apple-title">
<span class="title-main">
<img src="static/images/icon.png" alt="VisGym Logo">
VisGym
</span>
<br><span class="title-subtitle">Diverse, Customizable, Scalable Environments for Multimodal Agents</span>
</h1>
</div>
<div class="hero-split-container page-shell">
<!-- Left Panel: Information Console -->
<div class="hero-left-panel">
<div class="hero-content-wrapper">
<!-- Authors Block with Accent Line -->
<div class="authors-block-module">
<div class="publication-authors apple-authors">
<span class="author-block">
<a href="https://zwcolin.github.io/" target="_blank">Zirui Wang</a><sup>†</sup>,</span>
<span class="author-block">
<a href="https://www.junyi42.com" target="_blank">Junyi Zhang</a><sup>†</sup>,</span>
<span class="author-block">
<a href="https://gejiaxin.org/" target="_blank">Jiaxin Ge</a><sup>†</sup>,</span>
<span class="author-block">
<a href="https://tonylian.com/" target="_blank">Long Lian</a>,</span>
<span class="author-block">
<a href="https://max-fu.github.io/" target="_blank">Letian Fu</a>,</span>
<span class="author-block">
<a href="https://lisabdunlap.com/" target="_blank">Lisa Dunlap</a>,</span>
<span class="author-block">
<a href="https://goldberg.berkeley.edu/" target="_blank">Ken Goldberg</a>,</span>
<span class="author-block">
<a href="https://people.eecs.berkeley.edu/~xdwang/" target="_blank">Xudong Wang</a>,</span>
<span class="author-block">
<a href="https://people.eecs.berkeley.edu/~istoica/" target="_blank">Ion Stoica</a>,</span>
<span class="author-block">
<a href="https://dchan.cc/" target="_blank">David M. Chan</a>,</span>
<span class="author-block">
<a href="https://www.sewonmin.com/" target="_blank">Sewon Min</a>,</span>
<span class="author-block">
<a href="https://joeygonzalez.com/" target="_blank">Joseph E. Gonzalez</a>
</span>
</div>
<div class="publication-authors apple-affiliations">
<span class="eql-cntrb"><small><sup>†</sup>Indicates Equal Contribution</small></span>
<div class="affiliation-group">
<span class="author-block">UC Berkeley</span>
<div class="lab-logos" id="labLogosContainer">
<a href="https://sky.cs.berkeley.edu/" target="_blank" class="lab-logo" title="Sky Lab">
<img src="static/images/skylab.png" alt="Sky Lab" class="lab-logo-img">
</a>
<a href="https://bair.berkeley.edu/" target="_blank" class="lab-logo" title="BAIR Lab">
<img src="static/images/BAIR_Logo.png" alt="BAIR Lab" class="lab-logo-img">
</a>
</div>
</div>
</div>
</div>
<!-- Action Buttons (Desktop: shown here, Mobile: hidden and shown in mobile container) -->
<div class="publication-links apple-links">
<span class="link-block">
<a href="https://huggingface.co/VisGym" target="_blank"
class="action-button action-button-hf" title="Benchmark / Dataset / Models">
<span class="button-icon">
<img src="static/images/hf-logo-pirate.png" alt="Hugging Face" class="hf-logo-img">
</span>
<span class="button-text">Benchmark / Dataset / Models</span>
</a>
</span>
<span class="link-block-newline"></span>
<span class="link-block">
<a href="https://github.com/visgym/VIsGym" target="_blank"
class="action-button action-button-code" title="Code">
<span class="button-icon">
<i class="fab fa-github"></i>
</span>
<span class="button-text">Code</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2601.16973" target="_blank"
class="action-button action-button-arxiv" title="arXiv">
<span class="button-icon">
<i class="ai ai-arxiv"></i>
</span>
<span class="button-text">arXiv</span>
</a>
</span>
</div>
</div>
</div>
<!-- Right Panel: Visual Window (Mobile: shown after abstract, before buttons) -->
<div class="hero-right-panel">
<div class="apple-image-container">
<img src="static/images/teaser.png" alt="VisGym Teaser" class="apple-teaser-image">
<p class="teaser-description">VisGym consists of 17 diverse, long-horizon environments designed to systematically evaluate, diagnose, and train VLMs on visually interactive tasks. The agent must select each action conditioned on both its past actions and observation history.</p>
</div>
</div>
</div>
<!-- Mobile: Buttons container (shown after teaser image) -->
<div class="hero-buttons-mobile-container">
<div class="hero-content-wrapper">
<div class="publication-links apple-links">
<span class="link-block">
<a href="https://huggingface.co/VisGym" target="_blank"
class="action-button action-button-hf" title="Benchmark / Dataset / Models">
<span class="button-icon">
<img src="static/images/hf-logo-pirate.png" alt="Hugging Face" class="hf-logo-img">
</span>
<span class="button-text">Benchmark / Dataset / Models</span>
</a>
</span>
<span class="link-block-newline"></span>
<span class="link-block">
<a href="https://github.com/visgym/VIsGym" target="_blank"
class="action-button action-button-code" title="Code">
<span class="button-icon">
<i class="fab fa-github"></i>
</span>
<span class="button-text">Code</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2601.16973" target="_blank"
class="action-button action-button-arxiv" title="arXiv">
<span class="button-icon">
<i class="ai ai-arxiv"></i>
</span>
<span class="button-text">arXiv</span>
</a>
</span>
</div>
</div>
</div>
</section>
<!-- End Apple-style Hero -->
<!-- Paper abstract -->
<section class="section hero is-light">
<div class="container is-max-desktop page-shell">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths abstract-column">
<h2 class="section-main-title left-align">Abstract</h2>
<div class="content has-text-justified abstract-content">
<p>Modern Vision–Language Models (VLMs) remain poorly characterized in multi-step visual interactions, particularly in how they integrate perception, memory, and action over long horizons.</p>
<p>We introduce <span class="highlight-contribution">VisGym</span>, a gymnasium of <span class="highlight-contribution">17 environments</span> for <span class="highlight-contribution">evaluating and training VLMs</span>. The suite spans symbolic puzzles, real-image understanding, navigation, and manipulation, and provides flexible controls over difficulty, input representation, planning horizon, and feedback. We also provide multi-step solvers that generate structured demonstrations, enabling supervised finetuning.</p>
<p>Our evaluations show that all frontier models struggle in interactive settings, achieving low success rates in both the easy (<span class="stat-easy">46.6%</span>) and hard (<span class="stat-hard">26.0%</span>) configurations. Our experiments reveal notable limitations: models struggle to effectively leverage long context, performing worse with an unbounded history than with truncated windows. Furthermore, we find that several text-based symbolic tasks become substantially harder once rendered visually.</p>
<p>However, explicit goal observations, textual feedback, and exploratory demonstrations in partially observable or unknown-dynamics settings for supervised finetuning yield <span class="highlight-negative">consistent gains</span>, highlighting concrete failure modes and pathways for improving multi-step visual decision-making.</p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- AI Agent Replay Dashboard -->
<div class="dashboard-embed-container embed-shell">
<iframe
id="visgym-frame"
src="replay_dashboard.html"
title="AI Agent Replay Dashboard"
allowfullscreen
scrolling="yes"
class="dashboard-iframe embed-frame"
style="width: 100%; border: none;">
</iframe>
</div>
<!-- End AI Agent Replay Dashboard -->
<!-- Frontier VLMs Fail Section -->
<section class="frontier-fail-section">
<div class="container is-max-desktop page-shell">
<div class="frontier-fail-content">
<h2 class="section-main-title">Frontier VLMs Fail on VisGym</h2>
<p class="frontier-fail-description">
Even the best-performing frontier model, Gemini-3-Pro, achieves only <strong class="stat-easy">46.61%</strong> (Easy) and <strong class="stat-hard">26.00%</strong> (Hard) on VisGym, indicating that VisGym poses a significant challenge for existing models.
</p>
<div class="frontier-fail-image-single">
<div class="frontier-fail-image-container">
<img src="static/images/avg_performance.png" alt="Average Performance" loading="lazy">
</div>
</div>
</div>
</div>
</section>
<!-- End Frontier VLMs Fail Section -->
<!-- Key Insights Section -->
<section class="key-insights-section">
<div class="max-w-6xl mx-auto px-4 py-6 bg-white">
<h2 class="section-main-title">Control Study</h2>
</div>
<!-- Section 1: Teaching Agents to Explore Before Exploitation -->
<section class="w-full pt-8 pb-6 bg-white text-gray-800">
<div class="max-w-6xl mx-auto px-4">
<div class="text-center mb-5">
<h3 class="text-xl font-bold berkeley-blue mb-2" style="font-family: 'Space Grotesk', sans-serif;">Teaching Agents to Explore Before Exploitation</h3>
</div>
<!-- Core Thesis -->
<div class="mb-5 space-y-3">
<p class="text-gray-700 text-base leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Some trajectories are far more informative than others—especially those that reveal hidden state or disambiguate perception. We test whether inducing such information-revealing behaviors during supervised finetuning improves VLM decision-making.
</p>
<p class="text-gray-700 text-base leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Across tasks with unknown dynamics and partial observability, structured demonstrations that explicitly probe action–perception correspondence consistently outperform solve-only trajectories, improving both success and final accuracy.
</p>
<p class="text-gray-700 text-base leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Further finetuning on longer but unstructured demonstrations degrades performance, confirming that gains come from the informative structure of demonstrations—not their length or quantity.
</p>
</div>
<!-- Comparative Experiment -->
<div class="mb-5">
<!-- Comparison Tables Side by Side -->
<div class="flex flex-col md:flex-row gap-6">
<!-- First Comparison Table -->
<div class="flex-1 flex flex-col">
<div class="mb-2 px-2">
<h4 class="text-base font-semibold text-gray-800" style="font-family: 'Space Grotesk', sans-serif;">Matchstick Rotation (Unknown Dynamics)</h4>
</div>
<div class="overflow-x-auto">
<table class="w-full border-collapse border border-gray-300 rounded-lg overflow-hidden shadow-lg bg-white">
<thead class="bg-gray-100">
<tr>
<th class="px-4 py-3 text-left text-gray-800 font-bold">Comparison Dimension</th>
<th class="px-4 py-3 text-center text-red-600 font-bold">Baseline Demonstrations (Stochastic)</th>
<th class="px-4 py-3 text-center text-green-600 font-bold">Information-Revealing (Structured)</th>
</tr>
</thead>
<tbody class="text-gray-700">
<tr class="border-t border-gray-200">
<td class="px-4 py-3 font-semibold">Strategy</td>
<td class="px-4 py-3 text-center">Three stochastic moves toward the target</td>
<td class="px-4 py-3 text-center">Two unit-scale steps to probe action–perception correspondence</td>
</tr>
<tr class="border-t border-gray-200 bg-gray-50/50">
<td class="px-4 py-4 font-semibold text-lg">Success Rate</td>
<td class="px-4 py-4 text-center text-red-600 font-bold text-2xl">32.9%</td>
<td class="px-4 py-4 text-center text-green-600 font-bold text-2xl">70.0% (≈2.1×)</td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Second Comparison Table -->
<div class="flex-1 flex flex-col">
<div class="mb-2 px-2">
<h4 class="text-base font-semibold text-gray-800" style="font-family: 'Space Grotesk', sans-serif;">Mental Rotation 3D (Partial Observability)</h4>
</div>
<div class="overflow-x-auto">
<table class="w-full border-collapse border border-gray-300 rounded-lg overflow-hidden shadow-lg bg-white">
<thead class="bg-gray-100">
<tr>
<th class="px-4 py-3 text-left text-gray-800 font-bold">Comparison Dimension</th>
<th class="px-4 py-3 text-center text-red-600 font-bold">Baseline Demonstrations (Solve-Only)</th>
<th class="px-4 py-3 text-center text-green-600 font-bold">Information-Revealing (Explore-Then-Solve)</th>
</tr>
</thead>
<tbody class="text-gray-700">
<tr class="border-t border-gray-200">
<td class="px-4 py-3 font-semibold">Strategy</td>
<td class="px-4 py-3 text-center">Directly attempt to solve the task</td>
<td class="px-4 py-3 text-center">Explicit exploratory actions to reveal hidden state</td>
</tr>
<tr class="border-t border-gray-200 bg-gray-50/50">
<td class="px-4 py-3 font-semibold">State Coverage</td>
<td class="px-4 py-3 text-center">Partial, implicit</td>
<td class="px-4 py-3 text-center">Explicitly disambiguates latent variables</td>
</tr>
<tr class="border-t border-gray-200">
<td class="px-4 py-4 font-semibold text-lg">Success Rate</td>
<td class="px-4 py-4 text-center text-red-600 font-bold text-2xl">28.6%</td>
<td class="px-4 py-4 text-center text-green-600 font-bold text-2xl">62.4% (≈2.2×)</td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
<!-- Final Insight -->
<div class="mt-3 p-4 border-l-4 border-green-600 bg-green-50 rounded-lg">
<p class="text-base font-medium text-green-900" style="font-size: 1rem; line-height: 1.625;">
Supervised finetuning is most effective when demonstrations teach agents how to reveal state, not just what action to take.
</p>
</div>
</div>
</section>
<!-- Section 2: Better Eyes Or Better Brain -->
<section class="w-full py-6 bg-white">
<div class="max-w-6xl mx-auto px-4">
<div class="flex flex-col md:flex-row-reverse gap-6 items-center">
<div class="w-full md:w-1/2">
<div class="bg-white p-4 rounded-2xl shadow-xl border border-gray-200">
<img src="static/images/sft_ablations.png" alt="SFT Ablations: Vision vs LLM Gains" class="w-full h-auto object-contain" style="image-rendering: -webkit-optimize-contrast; image-rendering: crisp-edges;">
</div>
</div>
<div class="w-full md:w-1/2 space-y-3">
<h3 class="text-xl font-bold berkeley-blue mb-1" style="font-family: 'Space Grotesk', sans-serif;">Better Eyes Or Better Brain</h3>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
We decoupled the architecture to ask a simple question: does performance come from <b>better eyes</b> (the vision encoder) or a <b>better brain</b> (the LLM)?
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Our analysis shows that, for most interactive tasks, <b>temporal reasoning</b> is the dominant factor.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
While visual perception is necessary, the ability to <b>integrate history and plan over time</b> is what truly differentiates model performance.
</p>
<div class="p-4 bg-blue-50 rounded-lg border-l-4 border-blue-600 text-base text-blue-900 italic" style="font-size: 1rem; line-height: 1.625;">
"Temporal reasoning and history integration remain the primary bottlenecks for current VLMs."
</div>
</div>
</div>
</div>
</section>
<!-- Section 3: Stronger Base Model Generalizes Better -->
<section class="w-full py-6 bg-white">
<div class="max-w-6xl mx-auto px-4">
<div class="flex flex-col md:flex-row gap-6 items-center">
<div class="w-full md:w-1/2">
<div class="bg-white p-4 rounded-2xl shadow-xl border border-gray-200">
<img src="static/images/qwen_easy_vs_hard_scatter_and_bar.png" alt="Qwen Easy vs Hard Scatter and Bar" class="w-full h-auto object-contain" style="image-rendering: -webkit-optimize-contrast; image-rendering: crisp-edges;">
</div>
</div>
<div class="w-full md:w-1/2 space-y-3">
<h3 class="text-xl font-bold berkeley-blue mb-1" style="font-family: 'Space Grotesk', sans-serif;">Stronger Base Model Generalizes Better</h3>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Supervised finetuning is known to generalize poorly to task variants. We revisit this question for modern VLMs by finetuning <b>Qwen2.5-VL-7B and Qwen3-VL-8B</b> on the same training data and optimization setup, then evaluating on harder task variants.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
While both models perform similarly on the easy variants seen during training, <b>Qwen3-VL generalizes substantially better</b> to harder settings, nearly doubling the success rate on average. This shows that newer VLMs expand the generalization limits of supervised finetuning in multi-step visual decision-making.
</p>
</div>
</div>
</div>
</section>
<!-- Section Divider -->
<div class="max-w-6xl mx-auto px-4 py-6 bg-white">
<h2 class="section-main-title">Diagnosing Frontier Models with VisGym</h2>
</div>
<!-- Section: Providing Final Goal at Beginning -->
<section class="w-full py-6 bg-white">
<div class="max-w-6xl mx-auto px-4">
<div class="flex flex-col gap-6">
<div class="w-full space-y-3">
<h3 class="text-xl font-bold berkeley-blue mb-1" style="font-family: 'Space Grotesk', sans-serif;">Providing Final Goal at Beginning</h3>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Providing the final solution image upfront reframes these tasks from reasoning about goals to aligning observations with a known target, shifting difficulty toward visual perception and tool execution.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
We evaluate this effect on five tasks—<b>Patch Reassembly, Jigsaw, Colorization, Zoom-In Puzzle, and Matchstick Equation</b>—where constructing the goal state is non-trivial. We augment instructions with the ground-truth final observation <i>o<sub>gt</sub></i>.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Across tasks, performance improves substantially, indicating that imagining or constructing the target state is a key bottleneck. However, accuracy remains far from perfect, revealing additional limitations beyond reasoning—most notably <b>fine-grained visual perception and action execution</b>.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Unexpectedly, <b>GPT-5 and Gemini 2.5 Pro</b> underperform on Zoom-In Puzzle and Matchstick Equation when the goal image is provided, often terminating early despite visible mismatches. Follow-up tests attribute this failure to visual misjudgment rather than reasoning errors: when asked whether initial and goal images were identical, Gemini 2.5 Pro produced false positives 80% and 57% of the time on these tasks, compared to 18%, 2%, and 0% on Colorization, Jigsaw, and Patch Reassembly.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
These results show that <b>perception errors can negate—or even reverse—the expected benefits of explicit goal supervision</b>.
</p>
</div>
<div class="w-full">
<div class="bg-white p-4 rounded-2xl shadow-xl border border-gray-200">
<img src="static/images/4.4.png" alt="Providing Final Goal at Beginning" class="w-full h-auto object-contain" style="image-rendering: -webkit-optimize-contrast; image-rendering: crisp-edges;">
<p class="text-sm text-gray-500 italic mt-4" style="font-size: 0.875rem; line-height: 1.5;">
<span class="font-semibold">Effect of providing the final goal observation.</span> No Final Obs. and With Final Obs. indicate whether the goal image is available at episode start (mean ± s.e.).
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Section: Turns to Keep in Conversation History -->
<section class="w-full py-6 bg-white">
<div class="max-w-6xl mx-auto px-4">
<div class="flex flex-col gap-6">
<div class="w-full space-y-3">
<h3 class="text-xl font-bold berkeley-blue mb-1" style="font-family: 'Space Grotesk', sans-serif;">Turns to Keep in Conversation History</h3>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
While longer interaction histories provide useful environmental signals, they also introduce <b>redundant and stale information that can hurt performance</b>.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Across <b>Maze2D, Sliding Block, MuJoCo Fetch Reach, and Matchstick Rotation</b>, models perform best with a limited recent history, but degrade when given the full unbounded context.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
This shows that visual context helps multi-step decision-making only up to a point. Importantly, the effect is <b>task- and model-dependent</b>, including cases of <b>reverse scaling</b> where longer history consistently reduces performance.
</p>
</div>
<div class="w-full">
<div class="bg-white p-4 rounded-2xl shadow-xl border border-gray-200">
<img src="static/images/4.1.png" alt="Turns to Keep in Conversation History" class="w-full h-auto object-contain" style="image-rendering: -webkit-optimize-contrast; image-rendering: crisp-edges;">
<p class="text-sm text-gray-500 italic mt-4" style="font-size: 0.875rem; line-height: 1.5;">
<span class="font-semibold">Effect of truncating conversational context.</span> Settings 1, 2, 4, and ∞ retain increasing amounts of recent history, from the current turn only to the full history (mean ± s.e.).
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Section: Removal of Text-based Feedback -->
<section class="w-full py-6 bg-white">
<div class="max-w-6xl mx-auto px-4">
<div class="flex flex-col gap-6">
<div class="w-full space-y-3">
<h3 class="text-xl font-bold berkeley-blue mb-1" style="font-family: 'Space Grotesk', sans-serif;">Removal of Text-based Feedback</h3>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Humans can infer action consequences directly from visual changes, but <b>current VLMs cannot reliably do so</b>.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
Across four tasks—<b>Maze 3D, Maze 2D, Sliding Block, and Matchstick Equation</b>—removing textual feedback and relying only on visual state transitions leads to consistent performance drops.
</p>
<p class="text-base text-gray-700 leading-relaxed" style="font-size: 1rem; line-height: 1.625;">
This shows that VLMs struggle to <b>judge action validity from visual changes alone</b> and depend heavily on text-based feedback during visually interactive decision-making.
</p>
</div>
<div class="w-full">
<div class="bg-white p-4 rounded-2xl shadow-xl border border-gray-200">
<img src="static/images/4.3.png" alt="Removal of Text-based Feedback" class="w-full h-auto object-contain" style="image-rendering: -webkit-optimize-contrast; image-rendering: crisp-edges;">
<p class="text-sm text-gray-500 italic mt-4" style="font-size: 0.875rem; line-height: 1.5;">
<span class="font-semibold">Effect of text-based feedback.</span> Results with and without environment feedback (mean ± s.e.).
</p>
</div>
</div>
</div>