-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
886 lines (854 loc) · 46.4 KB
/
Copy pathindex.html
File metadata and controls
886 lines (854 loc) · 46.4 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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta
name="description"
content="LongPuzzleBench tests GUI agents on interactive visual puzzles where early decisions shape every move that remains. Explore results, trajectories, and playable games."
/>
<meta name="theme-color" content="#f2eee5" />
<link rel="canonical" href="https://azurestarz.github.io/LongPuzzleBench/" />
<title>LongPuzzleBench · Long-horizon visual puzzle agents</title>
<script>
(() => {
const params = new URLSearchParams(window.location.search);
if (!params.has("game")) return;
const target = new URL("play/", window.location.href);
target.search = window.location.search;
target.hash = window.location.hash;
window.location.replace(target.href);
})();
</script>
<link rel="preload" href="assets/games/bolt_unscrew.webp" as="image" />
<link rel="preload" href="assets/games/rush_hour_2.webp" as="image" />
<link rel="stylesheet" href="assets/site.css" />
<link rel="stylesheet" href="assets/home.css" />
<script src="assets/site.js" defer></script>
<script src="assets/home-data.js" defer></script>
<script src="assets/home.js" defer></script>
</head>
<body>
<a class="skip-link" href="#main-content">Skip to project overview</a>
<header class="global-header" data-menu-open="false">
<div class="site-container global-header__inner">
<a class="project-brand" href="./" aria-label="LongPuzzleBench home">
<span class="project-brand__mark" aria-hidden="true"></span>
<span class="project-brand__copy">
<strong>LongPuzzleBench</strong>
<small>Visual agent benchmark</small>
</span>
</a>
<nav class="global-nav" id="project-navigation" aria-label="Project navigation">
<a href="./" aria-current="page">Home</a>
<a href="research/">Research</a>
<a href="play/">Playground</a>
<a href="#benchmark">Benchmark</a>
<a href="#updates">Updates</a>
<a class="global-nav__mobile-external" href="https://github.com/AzureStarz/LongPuzzleBench">GitHub ↗</a>
</nav>
<div class="global-header__actions">
<a class="global-header__github" href="https://github.com/AzureStarz/LongPuzzleBench">
GitHub <span aria-hidden="true">↗</span>
</a>
<button
class="nav-toggle"
type="button"
aria-controls="project-navigation"
aria-expanded="false"
aria-label="Open project navigation"
data-nav-toggle
>
<span class="nav-toggle__lines" aria-hidden="true"></span>
</button>
</div>
</div>
</header>
<main id="main-content">
<section class="home-hero" aria-labelledby="home-title">
<div class="site-container home-hero__layout">
<div class="home-hero__copy">
<p class="home-kicker">Long-horizon visual puzzle environments</p>
<h1 id="home-title"><span>Long</span><span>PuzzleBench</span></h1>
<p class="home-hero__claim">
Interactive puzzles where one legal move can decide whether dozens of future moves
still exist.
</p>
<p class="home-hero__dek">
LongPuzzleBench evaluates GUI agents across six stateful browser games and 114
levels—then opens their trajectories to inspection.
</p>
<div class="home-hero__actions" aria-label="Primary project actions">
<a class="project-button project-button--primary" href="research/">
Explore research <span aria-hidden="true">→</span>
</a>
<a class="project-button project-button--secondary" href="play/">
Play the games <span aria-hidden="true">→</span>
</a>
</div>
<div class="home-hero__text-links">
<a href="https://github.com/AzureStarz/LongPuzzleBench">View on GitHub ↗</a>
<a href="https://github.com/AzureStarz/LongPuzzleBench#evaluation-protocol">Evaluation protocol ↗</a>
</div>
<dl class="home-hero__facts" aria-label="Benchmark scope">
<div><dt>Games</dt><dd>6</dd></div>
<div><dt>Levels</dt><dd>114</dd></div>
<div><dt>Evaluation cells</dt><dd>16</dd></div>
</dl>
</div>
<div class="hero-sequence" aria-label="Three representative LongPuzzleBench puzzle states">
<svg class="hero-sequence__path" viewBox="0 0 720 420" aria-hidden="true">
<path d="M52 225 C 150 118, 238 320, 344 211 S 542 104, 672 218" />
<circle cx="52" cy="225" r="5" />
<circle cx="344" cy="211" r="5" />
<circle cx="672" cy="218" r="5" />
</svg>
<figure class="hero-state hero-state--access">
<div class="hero-state__image">
<img
src="assets/games/bolt_unscrew.webp"
width="360"
height="640"
alt="Bolt Unscrew board with one plank, screws, and an open destination"
fetchpriority="high"
decoding="async"
/>
<span>01</span>
</div>
<figcaption><strong>Access</strong><small>What stays reachable?</small></figcaption>
</figure>
<figure class="hero-state hero-state--identity">
<div class="hero-state__image">
<img
src="assets/games/rush_hour_2.webp"
width="360"
height="640"
alt="Rush Hour board with independently moving vehicles"
fetchpriority="high"
decoding="async"
/>
<span>19</span>
</div>
<figcaption><strong>Identity</strong><small>Which object moved?</small></figcaption>
</figure>
<figure class="hero-state hero-state--space">
<div class="hero-state__image">
<img
src="assets/games/color_connect.webp"
width="360"
height="640"
alt="Color Connect board with colored endpoints sharing a small grid"
decoding="async"
/>
<span>42</span>
</div>
<figcaption><strong>Shared space</strong><small>What must remain free?</small></figcaption>
</figure>
<p class="hero-sequence__note"><span aria-hidden="true">↳</span> Same action space. A different future after every move.</p>
</div>
</div>
</section>
<section class="home-section updates-section" id="updates" aria-labelledby="updates-title">
<div class="site-container">
<div class="home-section-heading home-section-heading--compact reveal">
<div>
<p class="home-kicker">Recent updates</p>
<h2 id="updates-title">The project is moving.</h2>
</div>
<p>Four concrete releases from the benchmark’s first public week.</p>
</div>
<ol class="updates-list reveal">
<li>
<time datetime="2026-08-28">28 Aug 2026</time>
<h3><a href="play/">Browser playground released</a></h3>
<p>Thirteen curated levels across all six game families are now playable without an install.</p>
</li>
<li>
<time datetime="2026-08-28">28 Aug 2026</time>
<h3><a href="research/">Trajectory research note published</a></h3>
<p>Five recurring failures are reconstructed from real model actions and board states.</p>
</li>
<li>
<time datetime="2026-08-28">28 Aug 2026</time>
<h3><a href="https://github.com/AzureStarz/LongPuzzleBench/blob/main/leaderboard/results.json">Leaderboard refreshed ↗</a></h3>
<p>Eighteen complete configurations now cover every evaluation cell.</p>
</li>
<li>
<time datetime="2026-08-27">27 Aug 2026</time>
<h3><a href="https://github.com/AzureStarz/LongPuzzleBench">Initial benchmark release ↗</a></h3>
<p>Six deterministic browser puzzle environments and 114 levels were opened.</p>
</li>
</ol>
<a class="home-text-action" href="https://github.com/AzureStarz/LongPuzzleBench/commits/main/">
View project history <span aria-hidden="true">↗</span>
</a>
</div>
</section>
<section class="home-section benchmark-intro" id="what-it-tests" aria-labelledby="tests-title">
<div class="site-container benchmark-intro__layout">
<div class="benchmark-intro__lead reveal">
<p class="home-kicker">What LongPuzzleBench tests</p>
<h2 id="tests-title">A puzzle is not a question.</h2>
<p>
A static answer can be right once. Here, an agent must keep its account of the world
right after every click, drag, swipe, collision, and state change.
</p>
</div>
<ol class="property-list reveal">
<li>
<span>01</span>
<div><h3>Interactive</h3><p>The board changes after every action. Observation and planning repeat until the level ends.</p></div>
<small>observe → act → rebuild</small>
</li>
<li>
<span>02</span>
<div><h3>Stateful</h3><p>Objects, selected sources, stopping positions, occupied cells, and exposed holes must persist.</p></div>
<small>pixels + memory</small>
</li>
<li>
<span>03</span>
<div><h3>Consequential</h3><p>A move can be legal now and still cover the only hole or consume the only route needed later.</p></div>
<small>local move ≠ safe move</small>
</li>
</ol>
</div>
</section>
<section class="home-section games-section" id="games" aria-labelledby="games-title">
<div class="site-container">
<div class="home-section-heading reveal">
<div>
<p class="home-kicker">Explore the suite</p>
<h2 id="games-title">Small interfaces. Long consequences.</h2>
</div>
<p>Six game families expose different failures in access, identity, ordering, coverage, workspace, and shared space.</p>
</div>
<ul class="game-showcase reveal">
<li>
<a class="showcase-card" href="play/?game=bolt-unscrew&difficulty=hard&level=1" style="--game-accent: #a7442d">
<span class="showcase-card__image">
<img src="assets/games/bolt_unscrew.webp" width="360" height="640" alt="Bolt Unscrew hard puzzle board" loading="lazy" decoding="async" />
<span>Two taps</span>
</span>
<div class="showcase-card__body">
<small>Irreversible access</small>
<h3>Bolt Unscrew</h3>
<span>Move screws without letting physics cover every future hole.</span>
<b>Play hard 1 <i aria-hidden="true">→</i></b>
</div>
</a>
</li>
<li>
<a class="showcase-card" href="play/?game=rush-hour&difficulty=hard&level=6" style="--game-accent: #b56574">
<span class="showcase-card__image">
<img src="assets/games/rush_hour_2.webp" width="360" height="640" alt="Rush Hour hard puzzle board" loading="lazy" decoding="async" />
<span>Drag</span>
</span>
<div class="showcase-card__body">
<small>Object identity</small>
<h3>Rush Hour</h3>
<span>Preserve vehicle identity while the target temporarily moves away from the exit.</span>
<b>Play hard 6 <i aria-hidden="true">→</i></b>
</div>
</a>
</li>
<li>
<a class="showcase-card" href="play/?game=nut-and-bolt&difficulty=hard&level=2" style="--game-accent: #76528d">
<span class="showcase-card__image">
<img src="assets/games/nut_and_bolt.webp" width="360" height="640" alt="Nut and Bolt hard puzzle board" loading="lazy" decoding="async" />
<span>Two taps</span>
</span>
<div class="showcase-card__body">
<small>Temporary workspace</small>
<h3>Nut and Bolt</h3>
<span>Sort 56 nuts while preserving enough room for later transfers.</span>
<b>Play hard 2 <i aria-hidden="true">→</i></b>
</div>
</a>
</li>
<li>
<a class="showcase-card" href="play/?game=truck-escape&difficulty=default&level=5" style="--game-accent: #617d6d">
<span class="showcase-card__image">
<img src="assets/games/truck_escape.webp" width="360" height="640" alt="Truck Escape puzzle board with many vehicles" loading="lazy" decoding="async" />
<span>Tap</span>
</span>
<div class="showcase-card__body">
<small>Dependency ordering</small>
<h3>Truck Escape</h3>
<span>Clear a 50-truck field in an order that keeps each heading open.</span>
<b>Play level 5 <i aria-hidden="true">→</i></b>
</div>
</a>
</li>
<li>
<a class="showcase-card" href="play/?game=maze-paint&difficulty=hard&level=4" style="--game-accent: #2d6b8b">
<span class="showcase-card__image">
<img src="assets/games/maze_paint.webp" width="360" height="640" alt="Maze Paint hard puzzle board" loading="lazy" decoding="async" />
<span>Swipe</span>
</span>
<div class="showcase-card__body">
<small>Persistent localization</small>
<h3>Maze Paint</h3>
<span>Cover every cell when the ball can stop only at a wall or edge.</span>
<b>Play hard 4 <i aria-hidden="true">→</i></b>
</div>
</a>
</li>
<li>
<a class="showcase-card" href="play/?game=color-connect&difficulty=hard&level=1" style="--game-accent: #c3697f">
<span class="showcase-card__image">
<img src="assets/games/color_connect.webp" width="360" height="640" alt="Color Connect hard puzzle board" loading="lazy" decoding="async" />
<span>Click path</span>
</span>
<div class="showcase-card__body">
<small>Shared space</small>
<h3>Color Connect</h3>
<span>Route eight pairs without consuming another color’s only corridor.</span>
<b>Play hard 1 <i aria-hidden="true">→</i></b>
</div>
</a>
</li>
</ul>
<a class="project-button project-button--secondary games-section__action" href="play/">
Open the full playground <span aria-hidden="true">→</span>
</a>
</div>
</section>
<section class="home-section results-section" id="benchmark" aria-labelledby="benchmark-title">
<div class="site-container">
<div class="results-heading reveal">
<div>
<p class="home-kicker">Benchmark frontier · 18 complete configurations</p>
<h2 id="benchmark-title">The leading configuration clears just over half the score scale.</h2>
</div>
<div class="results-heading__scope" aria-label="Benchmark scope">
<span><strong>6</strong> games</span>
<span><strong>114</strong> levels</span>
<span><strong>16</strong> score cells</span>
</div>
</div>
<!-- LONGPUZZLEBENCH:RESULTS:START -->
<div class="results-layout">
<aside class="frontier-panel" aria-labelledby="frontier-title">
<div class="frontier-panel__intro">
<p class="home-kicker">Frontier by model ID</p>
<h3 id="frontier-title">Best observed setting per model</h3>
<p>Each card keeps only the strongest reasoning setting for that model ID. The canonical configuration ranking remains at right.</p>
</div>
<ol class="frontier-models">
<li data-model="gpt-5.6-sol" data-effort="medium">
<span class="frontier-model__rank">01</span>
<span class="frontier-model__mark" aria-hidden="true">SOL</span>
<span class="frontier-model__identity">
<strong>GPT-5.6 Sol</strong>
<small>medium reasoning · leader</small>
</span>
<span class="frontier-model__score">
<strong>54.796</strong><small>/100</small>
<em>55.42% success</em>
</span>
</li>
<li data-model="gpt-5.6-terra" data-effort="high">
<span class="frontier-model__rank">02</span>
<span class="frontier-model__mark" aria-hidden="true">TER</span>
<span class="frontier-model__identity">
<strong>GPT-5.6 Terra</strong>
<small>high reasoning · −7.205 vs. leader</small>
</span>
<span class="frontier-model__score">
<strong>47.591</strong><small>/100</small>
<em>48.75% success</em>
</span>
</li>
<li data-model="kimi-k3" data-effort="high">
<span class="frontier-model__rank">03</span>
<span class="frontier-model__mark" aria-hidden="true">K3</span>
<span class="frontier-model__identity">
<strong>Kimi K3</strong>
<small>high reasoning · −22.102 vs. leader</small>
</span>
<span class="frontier-model__score">
<strong>32.694</strong><small>/100</small>
<em>32.50% success</em>
</span>
</li>
</ol>
<div class="metric-note">
<span aria-hidden="true">↳</span>
<p><strong>Primary metric</strong> Equal-weight mean across 16 game × difficulty cells. Bars use the released 0–100 scale.</p>
</div>
<p class="baseline-note"><strong>Reference point</strong> Human baseline is not reported in this release.</p>
</aside>
<div class="ranking-panel">
<header class="ranking-panel__header">
<div>
<p class="home-kicker">Complete ranking</p>
<h3>All 18 model × reasoning configurations</h3>
</div>
<div class="ranking-legend" aria-label="Ranking columns"><span>Score</span><span>Success</span></div>
</header>
<ol class="ranking-list" data-ranking-list>
<li class="ranking-row" data-rank="1" data-model="gpt-5.6-sol" data-effort="medium" data-score="54.796" data-success="55.42">
<span class="ranking-row__rank">01</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Sol</strong>
<small>gpt-5.6-sol</small>
</span>
<span class="ranking-row__effort">medium</span>
<span class="ranking-row__bar" aria-label="Score 54.796 out of 100"><i style="--score: 54.796%"></i></span>
<strong class="ranking-row__score">54.796</strong>
<span class="ranking-row__success">55.42%</span>
</li>
<li class="ranking-row" data-rank="2" data-model="gpt-5.6-sol" data-effort="low" data-score="54.142" data-success="55.62">
<span class="ranking-row__rank">02</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Sol</strong>
<small>gpt-5.6-sol</small>
</span>
<span class="ranking-row__effort">low</span>
<span class="ranking-row__bar" aria-label="Score 54.142 out of 100"><i style="--score: 54.142%"></i></span>
<strong class="ranking-row__score">54.142</strong>
<span class="ranking-row__success">55.62%</span>
</li>
<li class="ranking-row" data-rank="3" data-model="gpt-5.6-sol" data-effort="high" data-score="49.968" data-success="50.47">
<span class="ranking-row__rank">03</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Sol</strong>
<small>gpt-5.6-sol</small>
</span>
<span class="ranking-row__effort">high</span>
<span class="ranking-row__bar" aria-label="Score 49.968 out of 100"><i style="--score: 49.968%"></i></span>
<strong class="ranking-row__score">49.968</strong>
<span class="ranking-row__success">50.47%</span>
</li>
<li class="ranking-row" data-rank="4" data-model="gpt-5.6-terra" data-effort="high" data-score="47.591" data-success="48.75">
<span class="ranking-row__rank">04</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Terra</strong>
<small>gpt-5.6-terra</small>
</span>
<span class="ranking-row__effort">high</span>
<span class="ranking-row__bar" aria-label="Score 47.591 out of 100"><i style="--score: 47.591%"></i></span>
<strong class="ranking-row__score">47.591</strong>
<span class="ranking-row__success">48.75%</span>
</li>
<li class="ranking-row" data-rank="5" data-model="gpt-5.6-terra" data-effort="medium" data-score="37.772" data-success="38.18">
<span class="ranking-row__rank">05</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Terra</strong>
<small>gpt-5.6-terra</small>
</span>
<span class="ranking-row__effort">medium</span>
<span class="ranking-row__bar" aria-label="Score 37.772 out of 100"><i style="--score: 37.772%"></i></span>
<strong class="ranking-row__score">37.772</strong>
<span class="ranking-row__success">38.18%</span>
</li>
<li class="ranking-row" data-rank="6" data-model="gpt-5.6-terra" data-effort="low" data-score="36.806" data-success="37.29">
<span class="ranking-row__rank">06</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Terra</strong>
<small>gpt-5.6-terra</small>
</span>
<span class="ranking-row__effort">low</span>
<span class="ranking-row__bar" aria-label="Score 36.806 out of 100"><i style="--score: 36.806%"></i></span>
<strong class="ranking-row__score">36.806</strong>
<span class="ranking-row__success">37.29%</span>
</li>
<li class="ranking-row" data-rank="7" data-model="kimi-k3" data-effort="high" data-score="32.694" data-success="32.50">
<span class="ranking-row__rank">07</span>
<span class="ranking-row__model">
<strong>Kimi K3</strong>
<small>kimi-k3</small>
</span>
<span class="ranking-row__effort">high</span>
<span class="ranking-row__bar" aria-label="Score 32.694 out of 100"><i style="--score: 32.694%"></i></span>
<strong class="ranking-row__score">32.694</strong>
<span class="ranking-row__success">32.50%</span>
</li>
<li class="ranking-row" data-rank="8" data-model="gpt-5.6-luna" data-effort="high" data-score="30.337" data-success="31.25">
<span class="ranking-row__rank">08</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Luna</strong>
<small>gpt-5.6-luna</small>
</span>
<span class="ranking-row__effort">high</span>
<span class="ranking-row__bar" aria-label="Score 30.337 out of 100"><i style="--score: 30.337%"></i></span>
<strong class="ranking-row__score">30.337</strong>
<span class="ranking-row__success">31.25%</span>
</li>
<li class="ranking-row" data-rank="9" data-model="gpt-5.6-luna" data-effort="medium" data-score="24.158" data-success="23.75">
<span class="ranking-row__rank">09</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Luna</strong>
<small>gpt-5.6-luna</small>
</span>
<span class="ranking-row__effort">medium</span>
<span class="ranking-row__bar" aria-label="Score 24.158 out of 100"><i style="--score: 24.158%"></i></span>
<strong class="ranking-row__score">24.158</strong>
<span class="ranking-row__success">23.75%</span>
</li>
<li class="ranking-row" data-rank="10" data-model="gpt-5.6-luna" data-effort="low" data-score="17.867" data-success="16.04">
<span class="ranking-row__rank">10</span>
<span class="ranking-row__model">
<strong>GPT-5.6 Luna</strong>
<small>gpt-5.6-luna</small>
</span>
<span class="ranking-row__effort">low</span>
<span class="ranking-row__bar" aria-label="Score 17.867 out of 100"><i style="--score: 17.867%"></i></span>
<strong class="ranking-row__score">17.867</strong>
<span class="ranking-row__success">16.04%</span>
</li>
<li class="ranking-row" data-rank="11" data-model="qwen/qwen3.8-27b" data-effort="default" data-score="16.915" data-success="16.25">
<span class="ranking-row__rank">11</span>
<span class="ranking-row__model">
<strong>Qwen 3.8 27B</strong>
<small>qwen/qwen3.8-27b</small>
</span>
<span class="ranking-row__effort">default</span>
<span class="ranking-row__bar" aria-label="Score 16.915 out of 100"><i style="--score: 16.915%"></i></span>
<strong class="ranking-row__score">16.915</strong>
<span class="ranking-row__success">16.25%</span>
</li>
<li class="ranking-row" data-rank="12" data-model="moonshotai/kimi-k2.5" data-effort="max" data-score="15.979" data-success="15.47">
<span class="ranking-row__rank">12</span>
<span class="ranking-row__model">
<strong>Kimi K2.5</strong>
<small>moonshotai/kimi-k2.5</small>
</span>
<span class="ranking-row__effort">max</span>
<span class="ranking-row__bar" aria-label="Score 15.979 out of 100"><i style="--score: 15.979%"></i></span>
<strong class="ranking-row__score">15.979</strong>
<span class="ranking-row__success">15.47%</span>
</li>
<li class="ranking-row" data-rank="13" data-model="qwen/qwen3.5-122b-a10b" data-effort="default" data-score="11.285" data-success="10.10">
<span class="ranking-row__rank">13</span>
<span class="ranking-row__model">
<strong>Qwen 3.5 122B A10B</strong>
<small>qwen/qwen3.5-122b-a10b</small>
</span>
<span class="ranking-row__effort">default</span>
<span class="ranking-row__bar" aria-label="Score 11.285 out of 100"><i style="--score: 11.285%"></i></span>
<strong class="ranking-row__score">11.285</strong>
<span class="ranking-row__success">10.10%</span>
</li>
<li class="ranking-row" data-rank="14" data-model="qwen/qwen3.6-35b-a3b" data-effort="default" data-score="6.580" data-success="5.68">
<span class="ranking-row__rank">14</span>
<span class="ranking-row__model">
<strong>Qwen 3.6 35B A3B</strong>
<small>qwen/qwen3.6-35b-a3b</small>
</span>
<span class="ranking-row__effort">default</span>
<span class="ranking-row__bar" aria-label="Score 6.580 out of 100"><i style="--score: 6.580%"></i></span>
<strong class="ranking-row__score">6.580</strong>
<span class="ranking-row__success">5.68%</span>
</li>
<li class="ranking-row" data-rank="15" data-model="qwen/qwen3.5-397b-a17b" data-effort="default" data-score="6.471" data-success="4.84">
<span class="ranking-row__rank">15</span>
<span class="ranking-row__model">
<strong>Qwen 3.5 397B A17B</strong>
<small>qwen/qwen3.5-397b-a17b</small>
</span>
<span class="ranking-row__effort">default</span>
<span class="ranking-row__bar" aria-label="Score 6.471 out of 100"><i style="--score: 6.471%"></i></span>
<strong class="ranking-row__score">6.471</strong>
<span class="ranking-row__success">4.84%</span>
</li>
<li class="ranking-row" data-rank="16" data-model="z-ai/glm-4.6v" data-effort="max" data-score="6.386" data-success="5.47">
<span class="ranking-row__rank">16</span>
<span class="ranking-row__model">
<strong>GLM-4.6V</strong>
<small>z-ai/glm-4.6v</small>
</span>
<span class="ranking-row__effort">max</span>
<span class="ranking-row__bar" aria-label="Score 6.386 out of 100"><i style="--score: 6.386%"></i></span>
<strong class="ranking-row__score">6.386</strong>
<span class="ranking-row__success">5.47%</span>
</li>
<li class="ranking-row" data-rank="17" data-model="qwen/qwen3-vl-235b-a22b-thinking" data-effort="default" data-score="5.314" data-success="3.59">
<span class="ranking-row__rank">17</span>
<span class="ranking-row__model">
<strong>Qwen3-VL 235B Thinking</strong>
<small>qwen/qwen3-vl-235b-a22b-thinking</small>
</span>
<span class="ranking-row__effort">default</span>
<span class="ranking-row__bar" aria-label="Score 5.314 out of 100"><i style="--score: 5.314%"></i></span>
<strong class="ranking-row__score">5.314</strong>
<span class="ranking-row__success">3.59%</span>
</li>
<li class="ranking-row" data-rank="18" data-model="qwen/qwen3-vl-30b-a3b-thinking" data-effort="default" data-score="5.134" data-success="4.43">
<span class="ranking-row__rank">18</span>
<span class="ranking-row__model">
<strong>Qwen3-VL 30B Thinking</strong>
<small>qwen/qwen3-vl-30b-a3b-thinking</small>
</span>
<span class="ranking-row__effort">default</span>
<span class="ranking-row__bar" aria-label="Score 5.134 out of 100"><i style="--score: 5.134%"></i></span>
<strong class="ranking-row__score">5.134</strong>
<span class="ranking-row__success">4.43%</span>
</li>
</ol>
</div>
</div>
<ol class="result-observations" aria-label="Three observations from the results">
<li><span>01</span><p><strong>The top three configurations are all Sol settings.</strong> Medium leads low by 0.654 points; high trails the leader by 4.828.</p></li>
<li><span>02</span><p><strong>Success alone does not set the rank.</strong> Sol-low solves 0.20 percentage points more levels, while Sol-medium earns the higher 16-cell score.</p></li>
<li><span>03</span><p><strong>No configuration leads every game.</strong> Sol-high leads Rush Hour; Terra-high leads Maze Paint and Color Connect; Kimi K3 leads Bolt Unscrew and Truck Escape.</p></li>
</ol>
<!-- LONGPUZZLEBENCH:RESULTS:END -->
<div class="result-links reveal">
<a class="project-button project-button--secondary" href="https://github.com/AzureStarz/LongPuzzleBench#leaderboard">
View evaluation details <span aria-hidden="true">↗</span>
</a>
<a class="home-text-action" href="https://github.com/AzureStarz/LongPuzzleBench/blob/main/leaderboard/results.json">
Download canonical results JSON <span aria-hidden="true">↗</span>
</a>
</div>
</div>
</section>
<section class="home-section trajectory-section" id="trajectories" aria-labelledby="trajectories-title">
<div class="site-container">
<div class="trajectory-section__heading reveal">
<div>
<p class="home-kicker">Scores tell us who leads. Trajectories show why.</p>
<h2 id="trajectories-title">The mistake often happens before the bad move.</h2>
</div>
<p>
Each case synchronizes real evaluator frames from two runs. Model text is copied from
the selected public action record; actions and evaluator outcomes remain separately labelled.
</p>
</div>
<ul class="trajectory-grammar reveal" aria-label="Trajectory evidence legend">
<li><span>01</span><strong>Game state</strong><small>What the screenshot shows</small></li>
<li><span>02</span><strong>Recorded output</strong><small>What the model emitted</small></li>
<li><span>03</span><strong>Executed action</strong><small>What the runtime performed</small></li>
<li><span>04</span><strong>Environment</strong><small>What changed next</small></li>
</ul>
<div class="trajectory-casebook reveal" data-casebook>
<div class="trajectory-case-tabs" role="tablist" aria-label="Representative trajectory cases">
<!-- LONGPUZZLEBENCH:CASE-TABS:START -->
<button type="button" role="tab" id="case-tab-rush" aria-controls="trajectory-case" aria-selected="true" tabindex="0" data-case-tab="rush"><span>01</span>Rush Hour</button>
<button type="button" role="tab" id="case-tab-bolt" aria-controls="trajectory-case" aria-selected="false" tabindex="-1" data-case-tab="bolt"><span>02</span>Bolt Unscrew</button>
<button type="button" role="tab" id="case-tab-color" aria-controls="trajectory-case" aria-selected="false" tabindex="-1" data-case-tab="color"><span>03</span>Color Connect</button>
<!-- LONGPUZZLEBENCH:CASE-TABS:END -->
</div>
<article
class="trajectory-case"
id="trajectory-case"
role="tabpanel"
aria-labelledby="case-tab-rush"
data-case-stage
>
<header class="trajectory-case__header">
<div class="trajectory-case__copy">
<p class="home-kicker" data-case-kicker>Case 01 · Part–whole segmentation</p>
<h3 data-case-title>One model sees a car. Another sees a fragment that is not there.</h3>
<p data-case-summary>
The same initial pixels produce the first meaningful divergence: one run moves a
real blocker; the other invents a vertical red vehicle and receives no visible effect.
</p>
</div>
<div class="trajectory-case__meta">
<strong data-case-stat>17.336-point benchmark gap · 3 actions versus 6</strong>
<span data-case-match>Byte-identical initial screenshot</span>
</div>
</header>
<div class="trajectory-branches" data-case-branches>
<!-- LONGPUZZLEBENCH:CASE-FALLBACK:START -->
<section class="trajectory-run trajectory-run--success" data-case-branch="0">
<header class="trajectory-run__header">
<div>
<span data-branch-label>Higher-ranked run</span>
<strong data-branch-model>GPT-5.6 Sol · low</strong>
</div>
<span class="trajectory-run__rank" data-branch-rank>Rank 2 · 54.142</span>
</header>
<figure class="trajectory-state">
<div class="trajectory-state__image">
<img src="research/assets/trajectories/rush_grouped-001-pre.webp" width="360" height="640" alt="The board contains one articulated horizontal target." loading="lazy" decoding="async" data-case-image />
<span data-case-state-label>Before action 01</span>
</div>
<figcaption data-branch-thesis>Treats the articulated sprite as one target</figcaption>
</figure>
<dl class="trajectory-record">
<div class="trajectory-record__output"><dt>Recorded model output</dt><dd data-case-output lang="und">Thought: 紫色竖车挡住红车所在横道，其下方同列一直到棋盘底部都有足够空间；先将它沿长轴向下拖到底，清空出口通道。</dd></div>
<div class="trajectory-record__action"><dt>Executed action</dt><dd><code data-case-action>drag((540,730) → (540,1250))</code></dd></div>
<div class="trajectory-record__environment"><dt>Environment</dt><dd><strong data-case-fact>game state: one vehicle · one motion axis</strong><span data-case-description>The board contains one articulated horizontal target.</span><small data-case-feedback>Public feedback · screen changed</small></dd></div>
</dl>
<p class="trajectory-run__outcome" data-branch-outcome>Success · 3 actions</p>
</section>
<section class="trajectory-run trajectory-run--warning" data-case-branch="1">
<header class="trajectory-run__header">
<div>
<span data-branch-label>Lower-ranked run</span>
<strong data-branch-model>GPT-5.6 Terra · low</strong>
</div>
<span class="trajectory-run__rank" data-branch-rank>Rank 6 · 36.806</span>
</header>
<figure class="trajectory-state">
<div class="trajectory-state__image">
<img src="research/assets/trajectories/rush_split-001-pre.webp" width="360" height="640" alt="The same sprite is split into a body and a ‘small red vertical car.’" loading="lazy" decoding="async" data-case-image />
<span data-case-state-label>Before action 01</span>
</div>
<figcaption data-branch-thesis>Splits one sprite into two vehicles</figcaption>
</figure>
<dl class="trajectory-record">
<div class="trajectory-record__output"><dt>Recorded model output</dt><dd data-case-output lang="und">Thought: Move the small red vertical vehicle above the target upward to clear the target row.</dd></div>
<div class="trajectory-record__action"><dt>Executed action</dt><dd><code data-case-action>drag((292,731) → (292,602))</code></dd></div>
<div class="trajectory-record__environment"><dt>Environment</dt><dd><strong data-case-fact>one sprite described as two vehicles</strong><span data-case-description>The same sprite is split into a body and a ‘small red vertical car.’</span><small data-case-feedback>Public feedback · no visible effect</small></dd></div>
</dl>
<p class="trajectory-run__outcome" data-branch-outcome>Success · 6 actions</p>
</section>
<!-- LONGPUZZLEBENCH:CASE-FALLBACK:END -->
</div>
<div class="trajectory-case__controls" aria-label="Trajectory controls">
<button type="button" data-case-prev><span aria-hidden="true">←</span> Previous</button>
<div class="trajectory-case__position">
<strong data-case-moment>Same pixels</strong>
<output data-case-output-count>01 / 04</output>
<div class="trajectory-case__dots" data-case-dots aria-label="Choose trajectory moment"></div>
</div>
<button type="button" data-case-toggle aria-pressed="true">Pause</button>
<button type="button" data-case-next>Next <span aria-hidden="true">→</span></button>
</div>
<footer class="trajectory-case__footer">
<p>
<span aria-hidden="true">↳</span>
Recorded text is shown in its original language. English case titles and takeaways are editorial analysis, not model quotes.
</p>
<div>
<a class="home-text-action" href="research/#finding-5" data-case-research>Read this finding <span aria-hidden="true">→</span></a>
<a class="project-button project-button--primary" href="play/?game=rush-hour&difficulty=medium&level=3" data-case-play>
Try this exact puzzle <span aria-hidden="true">→</span>
</a>
</div>
</footer>
</article>
</div>
<aside class="trajectory-research-bridge reveal" aria-labelledby="research-deep-title">
<div class="research-atlas" aria-hidden="true">
<figure><img src="research/assets/trajectories/bolt_immediate-010-post.webp" width="360" height="640" alt="" loading="lazy" decoding="async" /><span>01 · Access</span></figure>
<figure><img src="research/assets/trajectories/nut_repair-074-post.webp" width="360" height="640" alt="" loading="lazy" decoding="async" /><span>02 · State</span></figure>
<figure><img src="research/assets/trajectories/color_recreate-017-post.webp" width="360" height="640" alt="" loading="lazy" decoding="async" /><span>03 · Space</span></figure>
<figure><img src="research/assets/trajectories/maze_stale-034-post.webp" width="360" height="640" alt="" loading="lazy" decoding="async" /><span>04 · Position</span></figure>
<figure><img src="research/assets/trajectories/rush_split-001-post.webp" width="360" height="640" alt="" loading="lazy" decoding="async" /><span>05 · Identity</span></figure>
</div>
<div class="trajectory-research-bridge__copy">
<p class="home-kicker">Trajectory analysis · 782 executed episodes</p>
<h3 id="research-deep-title">The move was legal. The puzzle was already lost.</h3>
<p>
Five recurring boundaries appear across the public runs: future actionability,
postcondition repair, shared-space reservation, self-localization, and object identity.
In Bolt Unscrew, 13 of 18 complete runs end with no available hole.
</p>
<div>
<a class="project-button project-button--light" href="research/">Read the full analysis <span aria-hidden="true">→</span></a>
<a href="research/data/findings.json">Inspect the evidence bundle ↗</a>
</div>
</div>
</aside>
</div>
</section>
<section class="home-section play-invitation" aria-labelledby="play-title">
<div class="site-container play-invitation__layout">
<div class="play-window reveal" aria-label="Maze Paint playground preview">
<div class="play-window__bar">
<span><i aria-hidden="true"></i> Human playground</span>
<small>Runtime loads after launch</small>
</div>
<div class="play-window__body">
<img src="assets/games/maze_paint.webp" width="360" height="640" alt="Maze Paint playable puzzle preview" loading="lazy" decoding="async" />
<div>
<p class="home-kicker">Suggested first puzzle</p>
<h3>Maze Paint · Easy 1</h3>
<p>Seven optimal swipes. The ball cannot stop midway.</p>
<span>Mouse + touch</span>
</div>
</div>
</div>
<div class="play-invitation__copy reveal">
<p class="home-kicker">Playground</p>
<h2 id="play-title">Think the puzzles look easy? Try one.</h2>
<p>
The public playground uses the same checked-in game runtime and mechanics as the
benchmark, without evaluator state, agent time limits, or diagnostic hints.
</p>
<ul aria-label="Playground features">
<li>13 curated demo levels</li>
<li>All six game families</li>
<li>No install or account</li>
</ul>
<div class="play-invitation__actions">
<a class="project-button project-button--primary" href="play/?game=maze-paint&difficulty=easy&level=1">
Play LongPuzzleBench <span aria-hidden="true">→</span>
</a>
<a class="home-text-action" href="play/">Browse all games <span aria-hidden="true">→</span></a>
</div>
</div>
</div>
</section>
<section class="home-section resources-section" aria-labelledby="resources-title">
<div class="site-container">
<div class="home-section-heading home-section-heading--compact reveal">
<div>
<p class="home-kicker">Project resources</p>
<h2 id="resources-title">Inspect the artifact.</h2>
</div>
<p>Source, protocol, results, and trajectory evidence remain public and reproducible.</p>
</div>
<ul class="resource-grid reveal">
<li>
<a href="https://github.com/AzureStarz/LongPuzzleBench">
<span>01 · Source</span><strong>GitHub repository</strong><small>Harness, environment source, configs, and tests.</small><i aria-hidden="true">↗</i>
</a>
</li>
<li>
<a href="https://github.com/AzureStarz/LongPuzzleBench#evaluation-protocol">
<span>02 · Method</span><strong>Evaluation protocol</strong><small>Observation, action, scoring, and progression rules.</small><i aria-hidden="true">↗</i>
</a>
</li>
<li>
<a href="https://github.com/AzureStarz/LongPuzzleBench/blob/main/leaderboard/results.json">
<span>03 · Results</span><strong>Machine-readable leaderboard</strong><small>Per-game and per-cell scores for all complete runs.</small><i aria-hidden="true">↗</i>
</a>
</li>
<li>
<a href="research/data/findings.json">
<span>04 · Evidence</span><strong>Trajectory findings bundle</strong><small>Claims, source hashes, story frames, and annotations.</small><i aria-hidden="true">→</i>
</a>
</li>
</ul>
</div>
</section>
</main>
<footer class="global-footer">
<div class="site-container">
<div class="global-footer__grid">
<div class="global-footer__intro">
<a class="project-brand" href="./" aria-label="LongPuzzleBench home">
<span class="project-brand__mark" aria-hidden="true"></span>
<span class="project-brand__copy">
<strong>LongPuzzleBench</strong>
<small>Visual agent benchmark</small>
</span>
</a>
<p>Interactive visual puzzles for studying whether GUI agents can preserve a coherent plan across long, consequential trajectories.</p>
</div>
<nav class="global-footer__links" aria-label="Footer navigation">
<a href="research/" data-local>Research story</a>
<a href="play/" data-local>Playground</a>
<a href="#benchmark" data-local>Benchmark snapshot</a>
<a href="#updates" data-local>Recent updates</a>
<a href="https://github.com/AzureStarz/LongPuzzleBench">GitHub</a>
<a href="https://github.com/AzureStarz/LongPuzzleBench#evaluation-protocol">Evaluation protocol</a>
</nav>
</div>
<div class="global-footer__meta">
<p>
LongPuzzleBench-authored material is available under Apache-2.0. Imported game material
and runtime components remain subject to their respective notices.
</p>
<span><a href="notices/NOTICE">Notices</a> · 2026 public release</span>
</div>
</div>
</footer>
</body>
</html>