-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
1566 lines (1430 loc) · 65.9 KB
/
index.html
File metadata and controls
1566 lines (1430 loc) · 65.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
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=1.0">
<title>Can LLMs Draw with SGP? - SphereLab</title>
<link rel="stylesheet" href="styles.css">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css">
<style>
/* 页面特定样式 */
:root {
--primary-color-dark: #8a8a8a;
}
body {
font-family: 'Inter', -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, Cantarell, 'Open Sans', 'Helvetica Neue', sans-serif;
line-height: 1.6;
color: var(--text-color);
background-color: #fff;
margin: 0;
padding: 0;
}
.container {
max-width: 1000px;
margin: 0 auto;
padding: 0 20px;
}
.article-container {
max-width: 800px;
margin: 0 auto;
}
.header-nav {
display: flex;
align-items: center;
padding: 20px 0;
border-bottom: 1px solid #eaeaea;
margin-bottom: 40px;
}
.home-link {
display: flex;
align-items: center;
text-decoration: none;
color: var(--text-color);
font-weight: 500;
transition: color 0.2s;
}
.home-link:hover {
color: var(--primary-color);
}
.home-link svg {
width: 32px;
height: 18px;
margin-right: 8px;
}
.article-header {
text-align: center;
margin-bottom: 40px;
}
.date-tag {
display: flex;
align-items: center;
justify-content: center;
gap: 16px;
margin-bottom: 20px;
}
.date {
font-size: 14px;
color: var(--secondary-text-color);
}
.tag {
font-size: 14px;
color: var(--primary-color);
font-weight: 500;
}
.article-title {
font-size: 40px;
font-weight: 700;
margin-bottom: 20px;
line-height: 1.2;
}
.article-abstract {
font-size: 16px;
color: var(--secondary-text-color);
max-width: 600px;
margin: 0 auto 30px;
text-align: left;
background-color: #e9f4f8;
padding: 20px;
border-radius: 10px;
font-style: italic;
}
.project-links {
display: flex;
flex-wrap: wrap;
gap: 10px;
justify-content: center;
margin-bottom: 40px;
}
.project-link {
display: inline-flex;
align-items: center;
padding: 8px 16px;
border-radius: 4px;
font-size: 14px;
font-weight: 500;
text-decoration: none;
transition: all 0.2s ease;
}
.project-link i {
margin-right: 8px;
}
.project-link.github {
background-color: #24292e;
color: white;
}
.project-link.github:hover {
background-color: #1a1e22;
}
.project-link.huggingface {
background-color: #ee9a1b;
color: white;
}
.project-link.huggingface:hover {
background-color: #ce8f32;
}
.project-link.arxiv {
background-color: #b31b1b;
color: white;
}
.project-link.arxiv:hover {
background-color: #9c1717;
}
.article-content {
font-size: 17px;
line-height: 1.7;
}
.article-content p {
margin-bottom: 24px;
}
.article-section {
margin: 60px 0;
}
.article-content h2 {
font-size: 28px;
font-weight: 700;
margin: 40px 0 24px;
scroll-margin-top: 80px;
}
.note {
font-style: italic;
color: var(--secondary-text-color);
font-size: 16px;
margin-bottom: 20px;
}
.feature-buttons {
display: flex;
flex-wrap: wrap;
gap: 10px;
margin: 30px 0;
scroll-margin-top: 80px;
}
.feature-button {
background-color: var(--light-gray);
border: none;
border-radius: 20px;
padding: 10px 18px;
font-size: 15px;
font-weight: 500;
cursor: pointer;
transition: background-color 0.2s, color 0.2s;
}
.feature-button:hover {
background-color: #e5e5e5;
}
.feature-button.active {
background-color: var(--primary-color);
color: white;
}
.reasoning-process {
margin: 30px 0;
border: 1px solid #eaeaea;
border-radius: 8px;
padding: 20px;
background-color: #f9f9f9;
}
.reasoning-container {
min-height: 200px;
}
.reasoning-loading {
display: flex;
flex-direction: column;
align-items: center;
justify-content: center;
gap: 16px;
}
.loading-spinner {
width: 40px;
height: 40px;
border: 4px solid rgba(0, 0, 0, 0.1);
border-radius: 50%;
border-top-color: var(--primary-color);
animation: spin 1s linear infinite;
}
@keyframes spin {
0% { transform: rotate(0deg); }
100% { transform: rotate(360deg); }
}
/* Navigation links for sections */
.section-nav {
position: fixed;
left: calc(50vw - 600px); /* 靠近内容区，540=800/2+140，内容最大宽800px，nav宽约140px */
top: 50%;
transform: translateY(-50%);
display: flex;
flex-direction: column;
gap: 15px;
padding: 20px;
background: rgba(255, 255, 255, 0.95);
border-radius: 10px;
box-shadow: 0 2px 10px rgba(0, 0, 0, 0.1);
z-index: 1000;
}
.section-link {
font-size: 14px;
font-weight: 500;
color: var(--secondary-text-color);
text-decoration: none;
transition: color 0.2s;
padding: 5px 10px;
white-space: nowrap;
}
.section-link:hover,
.section-link.active {
color: var(--primary-color);
background: rgba(0, 0, 0, 0.05);
border-radius: 5px;
}
/* Subsection navigation styling */
.section-link[href^="#visual-search"],
.section-link[href^="#visual-puzzle"],
.section-link[href^="#geography"],
.section-link[href^="#medical"],
.section-link[href^="#science"],
.section-link[href^="#information"],
.section-link[href^="#more"] {
font-size: 12px;
padding-left: 20px;
color: #888;
}
/* 作者栏左右分栏布局 */
.authors-section {
display: flex;
justify-content: space-between;
align-items: flex-start;
gap: 40px;
background: #f3f3f3;
border-radius: 10px;
padding: 24px 24px 24px 24px;
margin: 20px 0 40px;
border-bottom: 1px solid #eaeaea;
}
.authors-meta {
flex: 1.3;
}
.authors-title, .affiliations-title, .published-title {
color: #a0a0a0;
font-size: 13px;
font-weight: 500;
letter-spacing: 1px;
margin-bottom: 8px;
text-transform: uppercase;
}
.authors-list {
display: block;
font-size: 16px;
font-weight: 400;
color: #111;
/* margin-bottom: 10px; */
}
.author {
font-size: 16px;
font-weight: 400;
}
.author sup {
font-size: 12px;
color: #a4a4a4;
font-weight: 600;
}
.author-note {
font-size: 13px;
color: #888;
margin-top: 8px;
font-style: normal;
text-align: left;
}
.authors-right {
flex: 1;
min-width: 180px;
text-align: left;
}
.affiliations-list {
font-size: 16px;
color: #111;
}
.published-date {
font-size: 16px;
color: #111;
}
@media (max-width: 768px) {
.authors-section {
flex-direction: column;
gap: 10px;
text-align: left;
}
.authors-right {
text-align: left;
min-width: 0;
}
}
/* 页脚样式 */
footer {
padding: 20px 0;
margin-top: 60px;
text-align: center;
}
.footer-content {
max-width: 600px;
margin: 0 auto;
}
.footer-social {
display: flex;
gap: 20px;
justify-content: center;
margin-bottom: 20px;
}
.social-icon {
display: inline-flex;
align-items: center;
justify-content: center;
width: 36px;
height: 36px;
border-radius: 50%;
color: #666;
transition: color 0.2s, transform 0.2s;
font-size: 18px;
background-color: #f5f5f5;
}
.social-icon:hover {
color: var(--primary-color);
transform: translateY(-3px);
background-color: #eeeeee;
}
.copyright {
font-size: 14px;
color: #666;
}
/* code style */
.bibtex-block {
background: #f7f7fa;
border: 1px solid #e0e0e0;
border-radius: 8px;
padding: 18px 18px 12px 18px;
margin: 32px 0 32px 0;
position: relative;
}
.bibtex-title {
font-size: 16px;
font-weight: 600;
color: #444;
margin-bottom: 10px;
letter-spacing: 1px;
}
.bibtex-pre {
font-size: 14px;
color: #333;
background: none;
border: none;
padding: 0;
margin: 0;
font-family: 'Fira Mono', 'Consolas', 'Menlo', monospace;
white-space: pre;
line-height: 1.6;
overflow-x: auto;
word-break: normal;
}
.bibtex-copy-btn {
position: absolute;
top: 18px;
right: 18px;
background: #ededed;
border: none;
border-radius: 4px;
padding: 4px 10px;
font-size: 13px;
color: #666;
cursor: pointer;
transition: background 0.2s;
}
.bibtex-copy-btn:hover {
background: #e0e0e0;
color: #222;
}
@media (max-width: 768px) {
.article-title {
font-size: 32px;
}
.feature-buttons {
justify-content: center;
}
.authors-list {
flex-direction: column;
gap: 10px;
}
body {
padding: 0 15px;
}
.container {
padding: 0;
}
.article-container {
padding: 0;
}
}
@media (max-width: 1200px) {
.section-nav {
position: static;
transform: none;
flex-direction: row;
flex-wrap: wrap;
justify-content: center;
margin: 30px 0;
padding: 15px 0;
border-top: 1px solid #eaeaea;
border-bottom: 1px solid #eaeaea;
background: none;
box-shadow: none;
left: 0;
}
.article-container {
padding-left: 0;
}
}
/* Reasoning process styling */
.reasoning-container {
max-height: 600px;
overflow-y: auto;
padding: 16px;
width: 100%;
}
.reasoning-error {
padding: 20px;
text-align: center;
color: #d32f2f;
display: flex;
flex-direction: column;
align-items: center;
}
.error-icon {
font-size: 24px;
margin-bottom: 10px;
}
.reasoning-message {
margin-bottom: 16px;
padding: 12px;
border-radius: 6px;
width: 100%;
box-sizing: border-box;
}
.reasoning-user {
background-color: #f1f5f9;
}
.reasoning-user strong {
display: block;
margin-bottom: 10px;
color: #1d75a5;
font-size: 1.1em;
}
.reasoning-assistant {
background-color: #ffffff;
box-shadow: 0 1px 3px rgba(0, 0, 0, 0.08);
}
.reasoning-text {
margin-bottom: 12px;
line-height: 1.5;
}
.reasoning-code {
margin: 12px 0;
border-radius: 6px;
overflow: hidden;
}
.reasoning-code pre {
margin: 0;
padding: 12px;
background-color: #2b2b2b;
color: #f8f8f2;
overflow-x: auto;
font-size: 14px;
line-height: 1.4;
}
.reasoning-interpreter-output {
margin: 12px 0;
padding: 12px;
background-color: #f7f7f9;
border-left: 4px solid #8a8a8a;
font-family: monospace;
white-space: pre-wrap;
}
.reasoning-image-wrap {
margin: 12px 0;
text-align: center;
}
.reasoning-image {
max-width: 100%;
max-height: 400px;
border-radius: 6px;
box-shadow: 0 2px 5px rgba(0, 0, 0, 0.1);
cursor: pointer;
transition: all 0.3s ease;
}
.reasoning-image.enlarged {
max-height: none;
max-width: 90%;
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.15);
z-index: 10;
position: relative;
}
.reasoning-answer {
margin-top: 16px;
}
.answer-box {
padding: 12px 16px;
background-color: #e8f5e9;
border: 1px solid #c8e6c9;
border-radius: 6px;
color: #388e3c;
font-weight: 600;
}
.user-question {
background-color: #e8f5fe;
border-left: 4px solid #1d75a5;
padding: 15px;
margin-bottom: 15px;
border-radius: 4px;
}
.user-question strong {
display: block;
color: #1d75a5;
margin-bottom: 8px;
font-size: 1.05em;
}
/* 对话气泡样式 */
.bubble {
max-width: 80%;
padding: 14px 18px;
border-radius: 18px;
font-size: 16px;
line-height: 1.6;
box-shadow: 0 2px 8px rgba(0,0,0,0.04);
word-break: break-word;
}
.bubble-gpt {
background: #ffffff;
color: #222;
align-self: flex-start;
border-top-left-radius: 6px;
margin-right: auto;
}
.bubble-human {
background: #f3f9ff;
color: #1a3a5d;
align-self: flex-end;
border-top-right-radius: 6px;
margin-left: auto;
}
.reasoning-message {
display: flex;
margin-bottom: 16px;
}
/* 新增：标识符样式 */
.tag-indicator {
font-family: 'Fira Mono', 'Consolas', 'Menlo', monospace;
font-size: 12px;
color: #666;
background-color: #f5f5f5;
padding: 2px 6px;
border-radius: 3px;
margin: 2px 0;
border: 1px solid #e0e0e0;
display: inline-block;
}
.reasoning-code-block,
.reasoning-image-block,
.reasoning-interpreter-block,
.reasoning-answer-block {
margin: 12px 0;
border: 1px solid #e8e8e8;
border-radius: 6px;
padding: 8px;
background-color: #fafafa;
}
/* Category sections styling */
.category-section {
margin: 50px 0;
border-bottom: 1px solid #eaeaea;
padding-bottom: 40px;
}
.category-section:last-child {
border-bottom: none;
}
.category-title {
font-size: 24px;
font-weight: 600;
margin-bottom: 20px;
color: var(--primary-color-dark);
}
.category-buttons {
margin-bottom: 30px;
}
</style>
</head>
<body>
<main class="article-container">
<article>
<div class="article-header" style="margin-top: 30px;">
<div class="date-tag">
<span class="date">September 9, 2025</span>
<span class="tag">Released</span>
</div>
<h1 class="article-title">Symbolic Graphics Programming with Large Language Models</h1>
<div style="text-align: center;margin-bottom: 30px;">
<img src="img/sgp-teaser.svg" alt="SVG tesaer"">
</div>
<!-- authors -->
<div class="authors-section">
<div class="authors-meta">
<div class="authors-title" style="text-align:left;">AUTHORS</div>
<div class="authors-list" style="text-align:left;">
<span class="author">Yamei Chen<sup>1,2,*</sup></span>,
<span class="author">Haoquan Zhang<sup>1,3,*</sup></span>,
<span class="author">Yangyi Huang<sup>1,2</sup></span>,
<span class="author">Zeju Qiu<sup>4</sup></span>,
<span class="author">Kaipeng Zhang<sup>3</sup></span>,
<span class="author">Yandong Wen<sup>2</sup></span>,
<span class="author">Weiyang Liu<sup>1,4</sup></span>
</div>
<div class="author-note">* Equal Contributing</div>
</div>
<div class="authors-right">
<div class="affiliations-title">AFFILIATIONS</div>
<div class="affiliations-list">
<sup>1</sup> CUHK<br>
<sup>2</sup> Westlake University<br>
<sup>3</sup> Shanghai AI Laboratory<br>
<sup>4</sup> MPI-IS
</div>
</div>
</div>
<div class="project-links">
<a href="https://github.com/Sphere-AI-Lab/SGP-RL" class="project-link github">
<i class="fab fa-github"></i> Code
</a>
<a href="https://huggingface.co/SphereLab/SGP-RL" class="project-link huggingface">
<i class="fas fa-rocket"></i> Model
</a>
<a href="https://huggingface.co/collections/SphereLab/sgp-generation-68b7d8a3b70ff81184c9923d" class="project-link huggingface">
<i class="fas fa-chart-bar"></i> Bench
</a>
<a href="https://arxiv.org/abs/2509.05208" class="project-link arxiv">
<i class="fas fa-file-alt"></i> Paper
</a>
</div>
</div>
<nav class="section-nav">
<a href="#introduction" class="section-link active">Introduction</a>
<a href="#sgp-gen-bench" class="section-link">SGP-Gen-Bench</a>
<a href="#method" class="section-link">Method</a>
<a href="#results" class="section-link">Results</a>
<a href="#object-scene" class="section-link" style="font-size: 12px; padding-left: 20px; color: #888;">Object & Scene</a>
<a href="#composition" class="section-link" style="font-size: 12px; padding-left: 20px; color: #888;">Composition</a>
<a href="#analysis" class="section-link">Analysis</a>
<a href="#bibtex" class="section-link">Bibtex</a>
</nav>
<div class="article-content">
<h2 id="introduction">Introduction</h2>
<p>
Large language models (LLMs) excel at understanding and generating code, but their ability to create symbolic graphics programs (SGPs)—which produce images remains underexplored. SVGs, as a widely used SGP format, provide an ideal testbed for evaluating LLMs in this area. We address two main research questions:
</p>
<p>
<span style="color:#1274c5; font-weight:bold;">(1) how well can LLMs draw using SGPs?</span>
<br>
<span style="color:#1274c5;">To address the first question, we introduce <em>SGP-GenBench</em>, a comprehensive benchmark for evaluating LLMs' ability to generate SGPs from three perspectives: object, scene, and composition.</span>
</p>
<p>
<span style="color:#970d0d; font-weight:bold;">(2) how can we improve their ability to generate SGPs?</span>
<br>
<span style="color:#970d0d;">To improve performance, we propose a reinforcement learning approach that uses similarity scores between visual encoder outputs and input text descriptions as reward signals. Our experiments show that RL significantly boosts the symbolic graphics programming abilities of LLMs.</span>
</p>
<p>
</p>
<h2 id="sgp-gen-bench">SGP as Visual Synthesis</h2>
<img src="img/parameterized- procedural.drawio.svg" alt="SVG visual synthesis"">
<p>
Symbolic graphics programming (SGP) aims to generate structured graphics code, such as SVG, from natural language instructions. Unlike conventional text-to-image models that produce pixel-based images, SGP translates prompts into interpretable, parametric programs, enabling precise and controllable visual synthesis.
</p>
<p>
SVG is widely used and serves as a natural bridge between language and vision, making it ideal for evaluating LLMs' ability to ground semantics from prompt to code.
SGPs are both:
</p>
<p>
<b style="color:#1274c5;">(a) PARAMETRIC</b>
<span style="color:#1274c5;"> - allowing exact geometric and appearance control</span>
<br>
<b style="color:#1ca67a;">(b) PROCEDURAL</b>
<span style="color:#1ca67a;"> - supporting hierarchical and compositional scene construction.</span>
</p>
<p>
These properties enable fine-grained, interpretable, and flexible visual generation, and allow us to directly analyze the complexity and structure of generated scenes.
</p>
<h2 id="sgp-gen-bench">SGP-GenBench</h2>
<h3> Construction </h3>
<p>
<ul>
<li><b>- Scene generation</b>: Evaluated on COCO-val, covering 80 object categories with complex scene descriptions. We sample 1,024 images for efficient and representative assessment.</li>
<li><b>- Object generation</b>: Assessed on SGP-Object-val, a set of 930 single-object examples with captions, focusing on the model's ability to render individual objects accurately.</li>
<li><b>- Compositional generation</b>: Tested on SGP-CompBench with 3,200 prompts, targeting attribute binding, spatial relationships, and numeracy using 80 common objects for compositional evaluation.</li>
</ul>
</p>
<h2 id="method">Method</h2>
<img src="img/method.drawio.svg" alt="Method" style="width:80%;max-width:1000px;display:block;margin:auto;">
<p>
Given a text description, we sample a group of SVG codes from the model and render them as images. Each SVG code is scored by the alignment between the rendered image and the text description. The advantages are calculated based on the scores, and used for updating the model.
</p>
<h3>Discussion of Method</h3>
<p></p>
<p>
<b>Distilling from vision models.</b>
We use external vision models (e.g., DINO, CLIP) to define rewards, guiding the LLM to align its generations with strong visual and semantic priors. This implicit distillation improves both visual understanding and cross-modal alignment.
</p>
<p>
<b>Training without ground truth SGPs.</b>
Our RL method learns directly from images using feedback from vision models, removing the need for paired image-program data and enabling scalable, diverse training.
</p>
<p>
<b>Aligning language and vision.</b>
The RL process ensures that generated SGPs are consistent with both linguistic input and visual model judgments, grounding language in visual evidence and enhancing cross-modal reasoning.
</p>
<h2 id="results">Results</h2>
<h3>Overview of SGP-GenBench</h3>
<p></p>
<!-- INSERT_YOUR_CODE -->
<ul style="line-height:1.7em;margin-bottom:18px;">
<li>
<b>SGP-GenBench reflects general model abilities.</b>
The ranking of models on our benchmark aligns well with their perceived general capabilities, especially in code generation. For example, Claude 3.7 Sonnet Thinking generally outperforms o3, which in turn surpasses Gemini 2.5 Pro Preview, followed by open-source systems like DeepSeek-R1 and Qwen-2.5-7B. This consistent ordering suggests that SVG generation is a reliable indicator of broader model competence.
</li>
<li>
<b>Closed-source models remain strongest.</b>
Frontier systems achieve the best results not only on compositional reasoning tasks such as attribute binding and numeracy, where Claude 3.7 Sonnet Thinking reaches 90.5 on color binding and 89.4 on numeracy, but also on scene and object fidelity, where Gemini 2.5 Pro Preview attains the top DINO object score of 0.653 and strong VQA scene performance of 0.554.
</li>
<li>
<b>Our RL-trained model substantially narrows the gap.</b>
The RL post-trained Qwen-2.5-7B raises its overall compositional score from 8.8 to 60.8, outperforming all other open-source counterparts such as DeepSeek-R1 and QwQ-32B. It also achieves the best VQA score across all models at 0.596, slightly higher than Claude 3.7 Sonnet Thinking, demonstrating that reinforcement learning enables open-source models to approach the closed-source frontier.
</li>
</ul>
<h3 id="object-scene">Object & Scene</h3>
<link rel="stylesheet" href="https://cdn.datatables.net/1.13.6/css/jquery.dataTables.min.css">
<style>
#sortable-results-table {
font-size: 14px;
background: #fff;
border: 1px solid #ddd;
margin: 20px 0;
width: 100%;
}
#sortable-results-table th, #sortable-results-table td {
border: 1px solid #ddd;
padding: 6px 8px;
text-align: center;
}
#sortable-results-table th {
background: #f8f9fa;
font-weight: 600;
cursor: pointer;
}
#sortable-results-table .group-header {
background: #f0f0f0;
font-weight: 600;
text-align: left;
}
#sortable-results-table .ours-row {
background: rgba(255, 165, 0, 0.15);
font-weight: 600;
}
</style>
<table id="sortable-results-table" style="width:100%">
<thead>
<tr>
<th rowspan="2" data-orderable="false">Model</th>
<th colspan="3">CLIP ↑</th>
<th colspan="3">DINO ↑</th>
<th colspan="3">VQA ↑</th>
<th colspan="3">HPS ↑</th>
</tr>
<tr>
<th>Sce.</th>
<th>Obj.</th>
<th>Avg.</th>
<th>Sce.</th>
<th>Obj.</th>
<th>Avg.</th>
<th>Sce.</th>
<th>Obj.</th>
<th>Avg.</th>
<th>Sce.</th>
<th>Obj.</th>
<th>Avg.</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="13">Frontier closed-source LLMs</td></tr>
<tr>
<td>GPT-4o-mini</td>
<td>0.207</td><td>0.278</td><td>0.243</td>
<td>0.021</td><td>0.573</td><td>0.297</td>
<td>0.295</td><td>0.465</td><td>0.380</td>
<td>0.118</td><td>0.174</td><td>0.146</td>
</tr>
<tr>
<td>GPT-4o</td>
<td>0.219</td><td>0.284</td><td>0.252</td>
<td>0.031</td><td>0.602</td><td>0.316</td>
<td>0.338</td><td>0.497</td><td>0.417</td>
<td>0.125</td><td>0.182</td><td>0.153</td>
</tr>
<tr>
<td>o1-mini</td>
<td>0.221</td><td>0.289</td><td>0.255</td>
<td>0.028</td><td>0.603</td><td>0.315</td>
<td>0.330</td><td>0.508</td><td>0.419</td>
<td>0.121</td><td>0.185</td><td>0.153</td>
</tr>
<tr>
<td>o1</td>
<td>0.220</td><td>0.285</td><td>0.252</td>
<td>0.031</td><td>0.607</td><td>0.319</td>
<td>0.354</td><td>0.520</td><td>0.437</td>
<td>0.122</td><td>0.185</td><td>0.153</td>
</tr>
<tr>
<td>o3-mini</td>
<td>0.231</td><td>0.293</td><td>0.262</td>
<td>0.036</td><td>0.608</td><td>0.322</td>
<td>0.379</td><td>0.520</td><td>0.450</td>
<td>0.128</td><td>0.187</td><td>0.158</td>
</tr>
<tr>
<td>o3</td>
<td>0.253</td><td>0.283</td><td>0.268</td>
<td>0.067</td><td>0.595</td><td>0.331</td>
<td>0.521</td><td>0.482</td><td>0.502</td>
<td>0.153</td><td>0.180</td><td>0.166</td>
</tr>
<tr>
<td>o4-mini</td>
<td>0.246</td><td>0.296</td><td>0.271</td>
<td>0.052</td><td>0.629</td><td>0.340</td>
<td>0.469</td><td>0.536</td><td>0.503</td>
<td>0.143</td><td>0.193</td><td>0.168</td>
</tr>
<tr>
<td>Gemini 2.0 Flash</td>
<td>0.204</td><td>0.275</td><td>0.240</td>
<td>0.023</td><td>0.588</td><td>0.306</td>
<td>0.293</td><td>0.468</td><td>0.381</td>
<td>0.116</td><td>0.176</td><td>0.146</td>
</tr>
<tr>
<td>Gemini 2.5 Flash Preview</td>
<td>0.222</td><td>0.286</td><td>0.254</td>
<td>0.033</td><td>0.603</td><td>0.318</td>
<td>0.349</td><td>0.498</td><td>0.424</td>
<td>0.125</td><td>0.183</td><td>0.154</td>
</tr>
<tr>
<td>Gemini 2.5 Pro Preview</td>
<td>0.256</td><td>0.302</td><td>0.279</td>
<td>0.088</td><td><b>0.653</b></td><td><b>0.371</b></td>
<td>0.554</td><td>0.572</td><td>0.563</td>
<td>0.154</td><td>0.199</td><td>0.177</td>
</tr>
<tr>
<td>Claude 3.5 Sonnet</td>
<td>0.240</td><td>0.293</td><td>0.266</td>
<td>0.055</td><td>0.624</td><td>0.340</td>
<td>0.428</td><td>0.528</td><td>0.478</td>
<td>0.140</td><td>0.190</td><td>0.165</td>
</tr>
<tr>
<td>Claude 3.7 Sonnet</td>
<td><b>0.262</b></td><td><b>0.306</b></td><td><b>0.284</b></td>
<td>0.088</td><td>0.647</td><td>0.368</td>
<td>0.581</td><td>0.567</td><td>0.574</td>
<td><b>0.165</b></td><td><b>0.200</b></td><td><b>0.183</b></td>
</tr>
<tr>
<td>Claude 3.7 Thinking</td>
<td><b>0.262</b></td><td>0.305</td><td><b>0.284</b></td>
<td><b>0.090</b></td><td>0.642</td><td>0.366</td>