-
Notifications
You must be signed in to change notification settings - Fork 6
Expand file tree
/
Copy pathindex.html
More file actions
946 lines (907 loc) · 40.5 KB
/
Copy pathindex.html
File metadata and controls
946 lines (907 loc) · 40.5 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Auto-Rubric as Reward</title>
<meta
name="description"
content="Auto-Rubric as Reward turns pairwise visual preferences into explicit, verifiable rubrics for text-to-image and image-editing optimization."
/>
<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=Fraunces:opsz,wght@9..144,500;9..144,700&family=Sora:wght@400;500;600;700;800&display=swap"
rel="stylesheet"
/>
<link rel="stylesheet" href="assets/site.css" />
</head>
<body>
<div class="page-shell">
<header class="site-header">
<div class="container header-row">
<a class="brand" href="#top" aria-label="Auto-Rubric as Reward home">
<span class="brand-copy">
<span class="brand-kicker">OpenEnvision</span>
<strong>Auto-Rubric as Reward</strong>
</span>
</a>
<nav class="site-nav" aria-label="Primary">
<a href="#overview">Overview</a>
<a href="#method">Method</a>
<a href="#benchmarks">Benchmarks</a>
<a href="#qualitative">Qualitative</a>
</nav>
</div>
</header>
<main id="top">
<section class="hero">
<div class="container hero-grid">
<div class="hero-copy" data-reveal>
<p class="eyebrow">Explicit reward criteria for visual generation</p>
<h1>Auto-Rubric as Reward</h1>
<p class="hero-subtitle">
From Implicit Preferences to Explicit Multimodal Generative
Criteria
</p>
<div class="author-block">
<p class="author-line">
<a
class="author-link"
href="https://tianshijing.github.io/"
target="_blank"
rel="noreferrer"
>Juanxi Tian<sup>1,2,*</sup></a>,
<a
class="author-link"
href="https://ferry1231.github.io/fengyuan/"
target="_blank"
rel="noreferrer"
>Fengyuan Liu<sup>1,*</sup></a>,
<a
class="author-link"
href="https://csuhan.com/"
target="_blank"
rel="noreferrer"
>Jiaming Han<sup>3</sup></a>,
<a
class="author-link"
href="https://scholar.google.com/citations?user=N4Du5-cAAAAJ&hl=zh-TW"
target="_blank"
rel="noreferrer"
>Yilei Jiang<sup>3</sup></a>,
<a
class="author-link"
href="https://yongliang-wu.github.io/"
target="_blank"
rel="noreferrer"
>Yongliang Wu<sup>4</sup></a>,
<a
class="author-link"
href="https://zhizhou57.github.io/"
target="_blank"
rel="noreferrer"
>Yesheng Liu<sup>1</sup></a>,
<a
class="author-link"
href="https://micky-li-hd.github.io/"
target="_blank"
rel="noreferrer"
>Haodong Li<sup>1</sup></a>,
<span class="author-link author-link-static">Furong Xu<sup>2</sup></span>,
<a
class="author-link"
href="https://li-wanhua.github.io/"
target="_blank"
rel="noreferrer"
>Wanhua Li<sup>1,†</sup></a>
</p>
<p class="affiliation-line">
<span><sup>1</sup>Nanyang Technological University</span>
<span><sup>2</sup>Ant Group</span>
<span><sup>3</sup>CUHK MMLab</span>
<span><sup>4</sup>UIUC</span>
</p>
<p class="author-note">
<span><sup>*</sup>Equal first authorship</span>
<span><sup>†</sup>Corresponding author</span>
</p>
</div>
<p class="hero-body">
Auto-Rubric as Reward converts a small set of labeled visual
supervision into readable rubric text, supports both pointwise
and pairwise VLM grading, and lets practitioners freely scale up
the rubric dimensions they care about. On top of that, we
provide a concise pairwise online RL algorithm for diffusion
models that emphasizes data efficiency, training stability, and
scalability, verifying that Rubric as Reward extends beyond
multimodal reasoning into multimodal generation, including
text-to-image and image editing.
</p>
<div class="hero-actions">
<a class="icon-button icon-button-paper is-disabled" href="#" aria-disabled="true">
<span class="button-icon" aria-hidden="true">
<img
src="https://arxiv.org/static/browse/0.3.4/images/arxiv-logomark-small-white.svg"
alt=""
width="22"
height="22"
/>
</span>
<span>Paper</span>
</a>
<a
class="icon-button icon-button-code"
href="https://github.com/OpenEnvision/AutoRubric-as-Reward"
target="_blank"
rel="noreferrer"
>
<span class="button-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" role="presentation">
<path
d="M12 .5C5.65.5.5 5.66.5 12.03c0 5.1 3.3 9.42 7.88 10.95.58.1.79-.25.79-.56 0-.28-.01-1.2-.02-2.17-3.2.7-3.88-1.36-3.88-1.36-.52-1.34-1.28-1.69-1.28-1.69-1.04-.72.08-.71.08-.71 1.15.08 1.76 1.19 1.76 1.19 1.02 1.77 2.67 1.26 3.33.96.1-.75.4-1.26.72-1.56-2.56-.29-5.24-1.29-5.24-5.73 0-1.27.45-2.31 1.18-3.12-.12-.29-.51-1.48.11-3.08 0 0 .97-.31 3.18 1.19a10.9 10.9 0 0 1 5.79 0c2.21-1.5 3.17-1.19 3.17-1.19.63 1.6.24 2.79.12 3.08.74.81 1.18 1.85 1.18 3.12 0 4.45-2.69 5.43-5.26 5.72.41.36.78 1.07.78 2.16 0 1.56-.02 2.82-.02 3.21 0 .31.2.67.8.56a11.55 11.55 0 0 0 7.87-10.95C23.5 5.66 18.35.5 12 .5Z"
/>
</svg>
</span>
<span>Code</span>
</a>
</div>
<div class="hero-tags" aria-label="Project scope">
<span>Text-to-Image</span>
<span>Image Editing</span>
<span>VLM-as-Judge</span>
<span>Auto-Rubric</span>
<span>Rubric as Reward</span>
</div>
</div>
<div class="hero-visual" data-reveal>
<div class="visual-stack card">
<div class="visual-kicker">
<span>Pipeline Overview</span>
<span>Final paper-aligned story</span>
</div>
<figure class="hero-figure">
<img
src="assets/site/figures/methods.png"
alt="Pipeline overview for Auto-Rubric as Reward"
/>
</figure>
</div>
<div class="floating-panel floating-panel-top">
<span class="panel-label">Core Thesis</span>
<strong>Readable rubrics replace hidden reward heuristics.</strong>
<p>
The judge no longer improvises a standard from scratch on every
comparison.
</p>
</div>
<div class="floating-panel floating-panel-bottom">
<span class="panel-label">Cross-Judge Benefit</span>
<strong>Explicit rubrics make judgment criteria portable.</strong>
<p>
Once the desired dimensions are written down, different VLM
judges stop improvising the evaluation standard and start
aligning around the same task-specific notion of quality.
</p>
</div>
</div>
</div>
</section>
<section class="thesis-strip" id="overview">
<div class="container thesis-grid" data-reveal>
<article class="thesis-card">
<span class="thesis-index">01</span>
<h2>Generate Criteria</h2>
<p>
Start from labeled supervision and ask a VLM to spell out the
exact visual dimensions that should matter for scoring or
ranking.
</p>
</article>
<article class="thesis-card">
<span class="thesis-index">02</span>
<h2>Verify Before Use</h2>
<p>
Keep only rubrics that recover the intended answer in pointwise
or pairwise settings, and revise them when they miss the target.
</p>
</article>
<article class="thesis-card">
<span class="thesis-index">03</span>
<h2>Auto-Rubric As Reward</h2>
<p>
Reuse the verified rubric set inside a frozen judge and connect
it to a minimal pairwise online RL loop for diffusion training.
</p>
</article>
</div>
</section>
<section class="section">
<div class="container intro-grid">
<div class="section-heading narrow" data-reveal>
<p class="section-kicker">Abstract</p>
<h2 class="section-title">Auto-Rubric as Reward</h2>
<p class="section-lead">
Aligning multimodal generative models with human preferences
requires reward signals that preserve the compositional and
multi-dimensional structure of judgment. Auto-Rubric as Reward
reframes reward modeling from implicit weight optimization to
explicit, criteria-based decomposition: before any comparison,
it externalizes a VLM's internalized preference knowledge into
prompt-specific rubrics, verifies those criteria against minimal
supervision, and consolidates them into a reusable structured
protocol for pointwise grading, pairwise evaluation, and reward
construction. By converting latent preference structure into
inspectable multimodal criteria, ARR reduces positional bias,
improves data efficiency, and exposes a stable factorized
interface for both zero-shot evaluation and downstream
generative alignment.
</p>
</div>
<div class="signal-cluster" data-reveal>
<article class="signal-panel card">
<span class="signal-tag">Why it matters</span>
<p>
The central bottleneck is not that VLMs lack preference
knowledge, but that existing scalar and pairwise objectives
fail to expose a stable factorized interface for applying it.
ARR addresses this mismatch by transforming holistic, latent
judgments into explicit and independently verifiable
multimodal criteria, thereby improving interpretability,
reducing reward hacking risk, and suppressing positional bias.
</p>
</article>
<article class="signal-panel card">
<span class="signal-tag">Where it works</span>
<p>
ARR supports both pointwise and pairwise VLM evaluation, then
extends naturally into multimodal generation through rubric-
conditioned policy optimization. In the paper, this interface
scales from evaluator fidelity benchmarks to text-to-image and
image-editing post-training, where explicit criteria become a
reusable supervision substrate rather than a task-specific
prompt trick.
</p>
</article>
</div>
</div>
</section>
<section class="section section-alt">
<div class="container">
<div class="section-heading" data-reveal>
<p class="section-kicker">Contributions</p>
<h2 class="section-title">Designed for interpretability without giving up performance</h2>
</div>
<div class="feature-grid">
<article class="feature-card card" data-reveal>
<span class="feature-number">A</span>
<h3>Explicit Reward Language</h3>
<p>
ARR externalizes implicit multimodal preferences into
prompt-conditioned natural-language rubrics that are
interpretable, verifiable, and highly data-efficient.
</p>
</article>
<article class="feature-card card" data-reveal>
<span class="feature-number">B</span>
<h3>Pointwise And Pairwise Auto-Rubric</h3>
<p>
The same rubric interface supports scalar-style pointwise
grading and pairwise comparison, allowing one structured
preference representation to unify evaluation, ranking, and
reward construction.
</p>
</article>
<article class="feature-card card" data-reveal>
<span class="feature-number">C</span>
<h3>Diagnosing the Interface Bottleneck</h3>
<p>
The paper argues that multimodal alignment is bottlenecked less
by missing knowledge than by the absence of a stable,
factorized interface for expressing and applying preference.
</p>
</article>
<article class="feature-card card" data-reveal>
<span class="feature-number">D</span>
<h3>A Concise Diffusion Online RL Algorithm</h3>
<p>
We also introduce a concise pairwise diffusion online RL
algorithm that emphasizes data efficiency, training stability,
and scalability, validating Rubric as Reward in text-to-image
and image editing rather than only multimodal reasoning.
</p>
</article>
</div>
</div>
</section>
<section class="section" id="method">
<div class="container">
<div class="section-heading" data-reveal>
<p class="section-kicker">Method</p>
<h2 class="section-title">From preference examples to reusable ARR rewards</h2>
</div>
<div class="method-grid">
<figure class="figure-card card figure-primary" data-reveal>
<a href="assets/image/methods.pdf">
<img
src="assets/site/figures/methods.png"
alt="Method diagram for Auto-Rubric as Reward"
/>
</a>
<figcaption>
The pipeline generates task-specific rubrics, verifies them on
supervision, organizes the surviving criteria into reusable
structure, and connects a frozen ARR judge to pairwise reward
construction for diffusion training.
</figcaption>
</figure>
<div class="method-notes">
<article class="note-card card" data-reveal>
<span class="note-index">Step 1</span>
<h3>Query-Specific Rubric Generation</h3>
<p>
Each supervised example becomes a prompt for extracting the
dimensions that should matter, whether the downstream task
needs pointwise scoring or pairwise ranking.
</p>
</article>
<article class="note-card card" data-reveal>
<span class="note-index">Step 2</span>
<h3>Verification And Revision</h3>
<p>
The same rubric must recover the desired supervision signal.
If the generated criteria fail under grading, the system
refines them instead of passing noisy reward logic
downstream.
</p>
</article>
<article class="note-card card" data-reveal>
<span class="note-index">Step 3</span>
<h3>Structured Reward Reuse</h3>
<p>
Verified rubrics are grouped into reusable themes and tips,
then consumed by a frozen VLM judge whose outputs can power
evaluation or pairwise online RL for generation models.
</p>
</article>
</div>
</div>
<div class="two-up rubric-showcase">
<figure class="figure-card card" data-reveal>
<a href="assets/image/arr_t2i.pdf">
<img
src="assets/site/figures/arr_t2i.png"
alt="Example text-to-image rubric produced by Auto-Rubric"
/>
</a>
<figcaption>
Text-to-image rubrics make scene composition, object
attributes, lighting realism, material quality, and artifact
control explicit.
</figcaption>
</figure>
<figure class="figure-card card" data-reveal>
<a href="assets/image/arr_edit.pdf">
<img
src="assets/site/figures/arr_edit.png"
alt="Example image-edit rubric produced by Auto-Rubric"
/>
</a>
<figcaption>
Editing rubrics emphasize instruction fulfillment, local edit
quality, preservation of source content, natural blending, and
avoidance of unintended changes.
</figcaption>
</figure>
</div>
</div>
</section>
<section class="section section-dark" id="benchmarks">
<div class="container">
<div class="section-heading section-heading-light" data-reveal>
<p class="section-kicker">Benchmarks</p>
<h2 class="section-title">Consistent gains across judges, preference benchmarks, and downstream generation quality</h2>
<p class="section-lead">
The paper shows two complementary effects: rubric-conditioning
makes VLM judging more reliable, and ARR-RPO translates that
better supervision into stronger text-to-image and image-editing
performance.
</p>
</div>
<div class="result-highlights">
<article class="result-stat card" data-reveal>
<span class="result-label">Pointwise / Pairwise</span>
<strong>One rubric interface</strong>
<p>Auto-Rubric can serve scalar grading, comparison, and reward construction in one workflow.</p>
</article>
<article class="result-stat card" data-reveal>
<span class="result-label">Custom Dimensions</span>
<strong>Scale what matters</strong>
<p>Researchers can expand rubric dimensions toward fidelity, preservation, composition, artifacts, or domain-specific constraints.</p>
</article>
<article class="result-stat card" data-reveal>
<span class="result-label">Generation Scope</span>
<strong>Beyond reasoning</strong>
<p>The paper validates Rubric as Reward on multimodal generation with diffusion-based text-to-image and image-editing training.</p>
</article>
</div>
<div class="results-grid">
<figure class="figure-card card results-main" data-reveal>
<a href="assets/image/performance.pdf">
<img
src="assets/site/figures/performance.png"
alt="Benchmark gains over specialist baseline models"
/>
</a>
<figcaption>
ARR-RPO improves FLUX.1-Dev and Qwen-Image-Edit on GenEval,
DPG-Bench, TIIF, UniGenBench++, GEdit-Bench, and ImgEdit.
</figcaption>
</figure>
<div class="results-side">
<figure class="figure-card card" data-reveal>
<a href="assets/image/cross_model.pdf">
<img
src="assets/site/figures/cross_model.png"
alt="Cross-model rubric transfer chart"
/>
</a>
<figcaption>
Structured rubrics transfer cleanly across judge families and
consistently outperform direct judging without explicit
criteria.
</figcaption>
</figure>
<article class="highlight-card card" data-reveal>
<h3>Reading the results</h3>
<ul>
<li>Explicit criteria improve alignment while keeping the reward interpretable and editable.</li>
<li>Pointwise and pairwise Auto-Rubric make the same rubric assets reusable across evaluation and training.</li>
<li>Custom rubric dimensions reduce reward drift toward generic aesthetic preference.</li>
<li>The diffusion RL recipe shows that rubric-conditioned rewards scale beyond multimodal reasoning tasks.</li>
</ul>
</article>
</div>
</div>
<div class="benchmark-panels">
<article class="benchmark-card card" data-reveal>
<div class="benchmark-head">
<div>
<p class="benchmark-kicker">Preference Evaluation</p>
<h3>Agreement with human labels across four preference benchmarks</h3>
</div>
<p class="benchmark-note">
Accuracy denotes how often the judge matches the annotated
preference.
</p>
</div>
<div class="table-wrap">
<table class="benchmark-table">
<thead>
<tr>
<th>Method</th>
<th>HPDv3</th>
<th>MM-RewardBench2 (T2I)</th>
<th>MM-RewardBench2 (Edit)</th>
<th>EditReward-Bench</th>
</tr>
</thead>
<tbody>
<tr class="group-row">
<td colspan="5">Trained reward model</td>
</tr>
<tr>
<td>PickScore</td>
<td>65.6</td>
<td>58.6</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>ImageReward</td>
<td>58.6</td>
<td>54.0</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>UnifiedReward</td>
<td>66.0</td>
<td>59.8</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>UnifiedReward-Thinking</td>
<td>68.1</td>
<td>66.0</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>HPSv3</td>
<td>76.9</td>
<td>60.2</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>EditReward</td>
<td>---</td>
<td>---</td>
<td>67.2</td>
<td>56.45</td>
</tr>
<tr class="group-row">
<td colspan="5">VLM-as-Judge (direct)</td>
</tr>
<tr>
<td>Qwen3-VL-8B</td>
<td>67.2</td>
<td>57.6</td>
<td>59.2</td>
<td>54.01</td>
</tr>
<tr>
<td>GPT-5</td>
<td>72.4</td>
<td>70.5</td>
<td>73.8</td>
<td>57.53</td>
</tr>
<tr>
<td>Gemini 3.1 Pro</td>
<td>76.6</td>
<td>75.1</td>
<td>77.4</td>
<td>61.23</td>
</tr>
<tr class="group-row arr-row">
<td colspan="5">ARR (ours)</td>
</tr>
<tr class="arr-row">
<td>Qwen3-VL-8B + ARR</td>
<td>70.2 <span class="gain">(+3.0)</span></td>
<td>62.7 <span class="gain">(+5.1)</span></td>
<td>65.5 <span class="gain">(+6.3)</span></td>
<td>57.22 <span class="gain">(+3.21)</span></td>
</tr>
<tr class="arr-row">
<td>GPT-5 + ARR</td>
<td>76.1 <span class="gain">(+3.7)</span></td>
<td>74.7 <span class="gain">(+4.2)</span></td>
<td>77.5 <span class="gain">(+3.7)</span></td>
<td>61.01 <span class="gain">(+3.48)</span></td>
</tr>
<tr class="arr-row best-row">
<td>Gemini 3.1 Pro + ARR</td>
<td>78.3 <span class="gain">(+1.7)</span></td>
<td>78.9 <span class="gain">(+3.8)</span></td>
<td>79.2 <span class="gain">(+1.8)</span></td>
<td>63.27 <span class="gain">(+2.04)</span></td>
</tr>
</tbody>
</table>
</div>
</article>
<article class="benchmark-card card" data-reveal>
<div class="benchmark-head">
<div>
<p class="benchmark-kicker">Generative Quality</p>
<h3>Text-to-image and image-editing gains under ARR-RPO</h3>
</div>
<p class="benchmark-note">
The strongest results come from rubric-conditioned judges,
especially with Gemini 3.1 Pro.
</p>
</div>
<div class="table-wrap">
<table class="benchmark-table benchmark-table-wide">
<thead>
<tr>
<th>Method</th>
<th>GenEval</th>
<th>DPG-Bench</th>
<th>TIIF</th>
<th>UniGen++ Short</th>
<th>UniGen++ Long</th>
<th>GEdit-Bench</th>
<th>ImgEdit</th>
</tr>
</thead>
<tbody>
<tr class="group-row">
<td colspan="8">Specialist model (T2I)</td>
</tr>
<tr>
<td>Emu3</td>
<td>0.54</td>
<td>80.60</td>
<td>--</td>
<td>45.42</td>
<td>50.59</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>JanusFlow</td>
<td>0.63</td>
<td>79.68</td>
<td>--</td>
<td>47.10</td>
<td>54.80</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>FLUX.1-Dev</td>
<td>0.66</td>
<td>83.84</td>
<td>71.09</td>
<td>60.97</td>
<td>69.42</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>DALL·E 3</td>
<td>0.67</td>
<td>83.50</td>
<td>74.96</td>
<td>68.85</td>
<td>70.82</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>Show-o2</td>
<td>0.76</td>
<td>86.14</td>
<td>--</td>
<td>61.90</td>
<td>70.33</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>OmniGen2</td>
<td>0.80</td>
<td>83.57</td>
<td>--</td>
<td>63.09</td>
<td>71.39</td>
<td>---</td>
<td>---</td>
</tr>
<tr>
<td>BAGEL</td>
<td>0.82</td>
<td>85.07</td>
<td>71.50</td>
<td>59.91</td>
<td>71.26</td>
<td>---</td>
<td>---</td>
</tr>
<tr class="group-row arr-row">
<td colspan="8">ARR-RPO / T2I (ours)</td>
</tr>
<tr class="arr-row">
<td>w/ RPO-Qwen3-VL-8B-ARR</td>
<td>0.74 <span class="gain">(+0.08)</span></td>
<td>85.03 <span class="gain">(+1.19)</span></td>
<td>74.92 <span class="gain">(+3.83)</span></td>
<td>64.17 <span class="gain">(+3.20)</span></td>
<td>71.82 <span class="gain">(+2.40)</span></td>
<td>---</td>
<td>---</td>
</tr>
<tr class="arr-row">
<td>w/ RPO-GPT-5-ARR</td>
<td>0.78 <span class="gain">(+0.12)</span></td>
<td>85.41 <span class="gain">(+1.57)</span></td>
<td>76.18 <span class="gain">(+5.09)</span></td>
<td>65.36 <span class="gain">(+4.39)</span></td>
<td>72.41 <span class="gain">(+2.99)</span></td>
<td>---</td>
<td>---</td>
</tr>
<tr class="arr-row best-row">
<td>w/ RPO-Gemini 3.1 Pro-ARR</td>
<td>0.80 <span class="gain">(+0.14)</span></td>
<td>85.76 <span class="gain">(+1.92)</span></td>
<td>76.85 <span class="gain">(+5.76)</span></td>
<td>65.89 <span class="gain">(+4.92)</span></td>
<td>72.93 <span class="gain">(+3.51)</span></td>
<td>---</td>
<td>---</td>
</tr>
<tr class="group-row">
<td colspan="8">Specialist model (editing)</td>
</tr>
<tr>
<td>Instruct-Pix2Pix</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>3.68</td>
<td>1.88</td>
</tr>
<tr>
<td>AnyEdit</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>3.21</td>
<td>2.45</td>
</tr>
<tr>
<td>Step1X-Edit</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>6.97</td>
<td>3.06</td>
</tr>
<tr>
<td>Qwen-Image-Edit-2509</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>7.54</td>
<td>4.35</td>
</tr>
<tr>
<td>UniWorldv2</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>7.76</td>
<td>4.48</td>
</tr>
<tr class="group-row arr-row">
<td colspan="8">ARR-RPO / image editing (ours)</td>
</tr>
<tr class="arr-row">
<td>w/ RPO-Qwen3-VL-8B-ARR</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>7.66 <span class="gain">(+0.12)</span></td>
<td>4.38 <span class="gain">(+0.03)</span></td>
</tr>
<tr class="arr-row">
<td>w/ RPO-GPT-5-ARR</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>7.72 <span class="gain">(+0.18)</span></td>
<td>4.40 <span class="gain">(+0.05)</span></td>
</tr>
<tr class="arr-row best-row">
<td>w/ RPO-Gemini 3.1 Pro-ARR</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>---</td>
<td>7.85 <span class="gain">(+0.31)</span></td>
<td>4.43 <span class="gain">(+0.08)</span></td>
</tr>
</tbody>
</table>
</div>
</article>
</div>
</div>
</section>
<section class="section" id="qualitative">
<div class="container">
<div class="section-heading" data-reveal>
<p class="section-kicker">Qualitative Results</p>
<h2 class="section-title">Stronger instruction fidelity, cleaner structures, and more faithful local edits</h2>
</div>
<div class="qualitative-grid">
<figure class="figure-card card example-top" data-reveal>
<a href="assets/image/results.pdf">
<img
src="assets/site/figures/results.png"
alt="Combined qualitative results for text-to-image and image editing"
/>
</a>
<figcaption>
ARR-RPO improves prompt satisfaction, structural plausibility,
and edit consistency while retaining source-image content when
preservation is part of the task.
</figcaption>
</figure>
<figure class="figure-card card" data-reveal>
<a href="assets/image/t2i_example.pdf">
<img
src="assets/site/figures/t2i_example.png"
alt="Text-to-image qualitative examples"
/>
</a>
<figcaption>
Text-to-image outputs show tighter content matching and more
deliberate compositional control under rubric-conditioned
optimization.
</figcaption>
</figure>
<figure class="figure-card card" data-reveal>
<a href="assets/image/edit_example.pdf">
<img
src="assets/site/figures/edit_example.png"
alt="Image-editing qualitative examples"
/>
</a>
<figcaption>
Image-editing outputs preserve identity and global scene
structure while executing localized instructions more
faithfully.
</figcaption>
</figure>
</div>
</div>
</section>
<section class="section citation-section">
<div class="container citation-grid">
<div data-reveal>
<p class="section-kicker">Citation</p>
<h2 class="section-title">BibTeX</h2>
<p class="section-lead">
If Auto-Rubric as Reward contributes to your work, please cite
the project and link back to the official repository.
</p>
</div>
<div class="code-card card" data-reveal>
<div class="code-card-head">
<span>Citation</span>
<span>arXiv link coming soon</span>
</div>
<pre><code>@misc{tian2026autorubricrewardimplicitpreferences,
title={Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria},
author={Juanxi Tian and Fengyuan Liu and Jiaming Han and Yilei Jiang and Yongliang Wu and Yesheng Liu and Haodong Li and Furong Xu and Wanhua Li},
year={2026},
eprint={2605.08354},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2605.08354},
}</code></pre>
</div>
</div>
</section>
</main>
<footer class="site-footer">
<div class="container footer-row">
<div>
<p class="footer-title">OpenEnvision</p>
<p class="footer-copy">
Project homepage for Auto-Rubric as Reward.
</p>
</div>
<div class="footer-links">
<a href="#top">Back to top</a>
<a href="https://github.com/OpenEnvision/AutoRubric-as-Reward" target="_blank" rel="noreferrer">Code</a>
<a href="#" aria-disabled="true">Paper</a>
</div>
</div>
</footer>
</div>
<script src="assets/site.js"></script>
</body>
</html>