-
Notifications
You must be signed in to change notification settings - Fork 1
/
index.html
704 lines (654 loc) · 35 KB
/
index.html
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Is Your Multimodal Language Model Oversensitive to Safe Queries?">
<meta name="keywords" content="MOSSBench, MOSS Bench">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>MOSSBench: Is Your Multimodal Language Model Oversensitive to Safe Queries?</title>
<!-- Global site tag (gtag.js) - Google Analytics -->
<!-- <script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script> -->
<!-- <script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
/Users/panlu/Library/Mobile Documents/com~apple~CloudDocs/ImageMath/visual-mathqa-server/data_final/images
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script> -->
<link rel="icon" href="./website/static/images/psych.webp">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="./website/static/css/bulma.min.css">
<link rel="stylesheet" href="./website/static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./website/static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./website/static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./website/static/css/index.css">
<link rel="stylesheet" href="./website/static/css/leaderboard.css">
<!-- <link href="https://unpkg.com/tabulator-tables@5.5.2/dist/css/tabulator_bulma.min.css" rel="stylesheet">
<script type="text/javascript" src="https://unpkg.com/tabulator-tables@5.5.2/dist/js/tabulator.min.js"></script> -->
<script type="text/javascript" src="website/static/js/sort-table.js" defer></script>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./website/static/js/fontawesome.all.min.js"></script>
<script src="./website/static/js/bulma-carousel.min.js"></script>
<script src="./website/static/js/bulma-slider.min.js"></script>
<script src="./website/static/js/explorer-index.js"></script>
<script src="./website/static/js/question_card.js"></script>
<script src="./website/static/js/leaderboard_testmini.js"></script>
<script src="./data/results/output_folders.js" defer></script>
<script src="./data/results/model_scores.js" defer></script>
<script src="./website/visualizer/data/data_public.js" defer></script>
</head>
<body>
<!-- <nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
<div class="navbar-item has-dropdown is-hoverable">
<a class="navbar-link">
More Research
</a>
<div class="navbar-dropdown">
<a class="navbar-item" href="https://ruocwang.github.io/">
<b>MoE</b> <p style="font-size:18px; display: inline; margin-left: 5px;">🔥</p>
</a>
<a class="navbar-item" href="https://xirui-li.github.io/DrAttack/">
<b>DrAttack</b> <p style="font-size:18px; display: inline; margin-left: 5px;">🔥</p>
</a>
<a class="navbar-item" href="https://ruocwang.github.io/">
<b>Negative Prompts</b> <p style="font-size:18px; display: inline; margin-left: 5px;">🔥</p>
</a>
<a class="navbar-item" href="https://measure-infinity.github.io/mulan/">
MuLan
</a>
</div>
</div>
</div>
</div>
</nav> --->
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title is-bold">
<img src="website/static/images/psych.webp" style="width:1em;vertical-align: middle" alt="Logo"/>
<span class="mathvista" style="vertical-align: middle">MOSSBench</span>
</h1>
<h2 class="subtitle is-3 publication-subtitle">
Is Your Multimodal Language Model Oversensitive to Safe Queries?
<!-- <br> -->
<!-- with GPT-4V, Bard, and Other Large Multimodal Models -->
</h2>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://xirui-li.github.io/">Xirui Li</a><sup style="color:#6fbf73;">1</sup><sup>*</sup>,</span>
<span class="author-block">
<a href="https://hengguangzhou.github.io/">Hengguang Zhou</a><sup style="color:#6fbf73;">1</sup><sup>*</sup>,</span>
<span class="author-block">
<a href="https://ruocwang.github.io/">Ruochen Wang</a><sup style="color:#6fbf73;">1</sup>,
</span>
<span class="author-block">
<a href="https://tianyizhou.github.io/">Tianyi Zhou</a><sup style="color:#ed4b82">2</sup>,
</span>
<span class="author-block">
<a href="https://cmhcbb.github.io/">Minhao Cheng</a><sup style="color:#ffac33">3</sup>,
</span>
<span class="author-block">
<a href="https://web.cs.ucla.edu/~chohsieh/">Cho-Jui Hsieh</a><sup style="color:#6fbf73;">1</sup>,
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup style="color:#6fbf73;">1</sup>University of California, Los Angeles,</span><br>
<span class="author-block"><sup style="color:#ed4b82">2</sup>University of Maryland, College Park,</span>
<span class="author-block"><sup style="color:#ffac33">3</sup>The Pennsylvania State University</span><br>
<span class="author-block">* Equal contribution</span><br>
<span class="paper-block">
<a href="https://turningpoint-ai.com" target="_blank" rel="external">
<img class="center-block org-banner" src="website/static/images/TurningPoint_transparent_cropped.png" style="height:10em">
</a>
</span>
</div>
<!-- <div class="is-size-5 publication-authors">
<span class="author-block">
<strong><span style="font-size: 1.5em; color: crimson;">A</span></strong>IGC
<strong><span style="font-size: 1.5em; color: crimson;">R</span></strong>esearch
<strong><span style="font-size: 1.5em; color: crimson;">C</span></strong>ollaboration
</span>
</div> -->
<!-- <div class="section" id="org-banners" style="display:fle">
<a href="https://turningpoint-ai.com" target="_blank" rel="external">
<img class="center-block org-banner" src="website/static/images/TurningPoint_transparent_cropped.png" style="height:15em">
</a>
</div> -->
<!-- <section> -->
<!-- <div class="section" id="org-banners" style="display:fle">
<a href="https://www.ucla.edu/" target="_blank" rel="external">
<img class="center-block org-banner" src="website/static/images/ucla.png" style="height:3em">
</a>
<a href="https://www.washington.edu/" target="blank" class="ext-link">
<img class="center-block org-banner" src="website/static/images/uw.png" style="height:3em">
</a>
<a href="https://www.microsoft.com/en-us/research/" target="_blank" rel="external">
<img class="center-block org-banner" src="website/static/images/microsoft.png" style="height:3em">
</a>
</div> -->
<!-- </section> -->
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<!-- @PAN TODO: change links -->
<a href="https://arxiv.org/pdf/2406.17806.pdf"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2406.17806"
class="external-link button is-normal is-rounded is-dark">
<!-- <a href="https://lupantech.github.io/papers/arxiv23_mathvista.pdf"
class="external-link button is-normal is-rounded is-dark"> -->
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Video Link. -->
<!-- <span class="link-block">
<a href="https://www.youtube.com/watch?v=MrKrnHhk8IA"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-youtube"></i>
</span>
<span>Video</span>
</a>
</span> -->
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/xirui-li/MOSSBench"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- Dataset Link. -->
<span class="link-block">
<a href="https://huggingface.co/datasets/AIcell/MOSSBench"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<!-- <i class="far fa-images"></i> -->
<p style="font-size:18px">🤗</p>
<!-- 🔗 -->
</span>
<span>Dataset</span>
</a>
</span>
<!-- Visualization Link. -->
<span class="link-block">
<a href="https://xirui-li.github.io/MOSSBench/#visualization"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<p style="font-size:18px">🔮</p>
</span>
<span>Visualize</span>
</a>
</span>
<!-- Leaderboard Link. -->
<!-- <span class="link-block">
<a href="https://xirui-li.github.io/MOSSBench/#leaderboard"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<p style="font-size:18px">🏆</p>
</span>
<span>Leaderboard</span>
</a>
</span> -->
<!-- Twitter Link. -->
<span class="link-block">
<a href="https://x.com/TurningPointAI/status/1807169697571553574"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<!-- <i class="far fa-images"></i> -->
<!-- 💻🔗 -->
<p style="font-size:18px">🌐</p>
</span>
<span>Twitter</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- <section class="hero teaser">
<div class="container is-max-desktop">
<div class="content has-text-centered">
<img src="website/static/images/tease_scores_gpt4v.png" alt="geometric reasoning" width="99%"/>
<p> Accuracy scores of one leading LLM (i.e., PoT GPT-4), four primary LMMs, random chance, and human performance our proposed
<img src="website/static/images/mathvista.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MathVista</span>
across mathematical reasoning and visual context types. PoT refers to program-of-thought prompting, and PoT GPT-4 is a textual LLM augmented with the caption and OCR text. GPT-4V is manually evaluated via the playground chatbot.
</p>
</div>
</div>
</section> -->
<section class="section">
<div class="container" style="margin-top: -150px; margin-bottom: -100px;">
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/main_plot.png" alt="geometric reasoning" width="100%"/>
<p><b>(Left)</b> MLLMs exhibit behaviors similar to human cognitive distortions, leading to <b class="best-score-text" style="color: #C6011F">oversensitive responses</b> where benign queries are perceived as harmful. We discover that oversensitivity prevails among existing MLLMs on
<img src="website/static/images/psych.webp" style="width:1.0em; vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span>.
</p>
<p><b>(Right)</b> Compliance rate of SOTA MLLMs on
<img src="website/static/images/psych.webp" style="width:1.0em; vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span>. Proprietary MLLMs (e.g., Claude 3, Gemini) exhibit more oversensitive behaviors on our dataset.
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container" style="margin-bottom: 2vh;">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Introduction</h2>
<div class="content has-text-justified">
<p>
Humans are prone to cognitive distortions — biased thinking patterns that lead to exaggerated responses to specific stimuli, albeit in very different contexts. This paper demonstrates that advanced <b>MLLMs</b> exhibit similar tendencies.
While these models are designed to respond queries under safety mechanism, they sometimes reject harmless queries in the presence of certain visual stimuli, disregarding the benign nature of their contexts.
</p>
<p>
As the initial step in investigating this behavior, we identify three types of stimulus that trigger the oversensitivity of existing MLLMs: <strong><em>Exaggerated Risk</em></strong>, <strong><em>Negated Harm</em></strong>, and <strong><em>Counterintuitive Interpretation</em></strong>.
To systematically evaluate MLLMs' oversensitivity to these stimuli, we propose the <b>M</b>ultimodal <b>O</b>ver<b>S</b>en<b>S</b>itivity <b>B</b>enchmark <img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
(<span class="mathvista">MOSSBench</span>). This toolkit consists of 300 manually collected benign multimodal queries, cross-verified by third-party reviewers (AMT).
</p>
<p>
Empirical studies using <img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span> on 20 MLLMs reveal several insights:
(1). Oversensitivity is prevalent among SOTA MLLMs, with refusal rates reaching up to 76% for harmless queries.
(2). Safer models are more oversensitive: increasing safety may inadvertently raise caution and conservatism in the model's responses.
(3). Different types of stimuli tend to cause errors at specific stages — perception, intent reasoning, and safety decision-making — in the response process of MLLMs.
These findings highlight the need for refined safety mechanisms that balance caution with contextually appropriate responses, improving the reliability of MLLMs in real-world applications.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<!-- Leaderboard -->
<!-- Oversensitivity Stimuli -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title is-1 mathvista">Oversensitivity Stimuli</h1>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<!-- <div class="column is-full-width has-text-centered"> -->
<div class="column is-four-fifths">
<h2 class="title is-3">Visual stimuli of model oversensitivity</h2>
<div class="content has-text-justified">
<p>
Through empirical investigation, we concluded three types of visual stimuli that trigger the oversensitivity of existing MLLMs: <strong><em>Exaggerated Risk</em></strong>, <strong><em>Negated Harm</em></strong>, and <strong><em>Counterintuitive Interpretation</em></strong>.
</p>
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/exaggerated_risk.png" alt="algebraic reasoning" width="40%"/>
<p>
Images in the wild, especially those with high resolution, are often rich in visual information. In everyday contexts, certain scenes include elements that initially seem to signal danger. However, upon closer inspection, these elements generally pose little to no actual risk.
Such a scenario frequently occurs in everyday life. MLLMs frequently refuse to process user requests involving these images, drastically overestimating their risk.
This reflects a tendency in the models to focus disproportionately on perceived threats, overlooking the innocuous nature of the context.
This behavior mirrors "catastrophizing," a cognitive distortion observed in human disorders.
We define these instances as <strong><em>Exaggerated Risk</em></strong>, where images containing safety-alerting visual elements carry minimal risk to the user's request.
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/negated_harm.png" alt="arithmetic reasoning" width="40%"/>
<p>
The second scenario, which we term <strong><em>Negated Harm</em></strong>, involves images where harmful objects or behaviors are present, but the overall context of the image actively opposes them.
This image depicts a prohibition sign over a scene where a man is physically threatening a woman.
Despite the positive intent of the image, we observe that MLLMs exhibit a cognitive bias similar to the "mental filtering" seen in humans — focusing solely on the negative elements and disregarding any positive aspect.
Specifically, the model overlooks the prohibition sign's context and concentrates only on the underlying scene of domestic violence.
Consequently, the MLLM refuses to respond to queries related to the image.
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/counterintuitive_interpretation.png" alt="arithmetic reasoning" width="40%"/>
<p>
The image features a girl holding a colorful parrot in a whimsical, fairy-tale setting.
The user inquires whether it is safe to put it in a cage.
Common sense would suggest that the query refers to the parrot.
However, our findings indicate that MLLMs often misinterpret such queries, falsely assuming the question concerns the safety of placing the girl in the cage instead.
This interpretation contradicts common human intuition and is highly unlikely to occur, a pattern we identify as <strong><em>Counterintuitive Interpretation</em></strong>.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- DATASET SECTION -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title is-1 mathvista">
<img src="website/static/images/psych.webp" style="width:1em;vertical-align: middle" alt="Logo"/>
<span class="mathvista" style="vertical-align: middle">MOSSBench Dataset</span>
</h1>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<!-- <div class="column is-full-width has-text-centered"> -->
<div class="column is-four-fifths">
<h2 class="title is-3">Overview</h2>
<div class="content has-text-justified">
<p>
<img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span> is the first benchmark for evaluating the oversensitivity of MLLMs systematically.
It consists of <b>300</b> samples, with different scenarios that contain <strong><em>Exaggerated Risk</em></strong>, <strong><em>Negated Harm</em></strong>, and <strong><em>Counterintuitive Interpretation</em></strong> as visual stimuli.
</p>
<p>
Collecting oversensitivity samples for MLLMs is challenging due to the intricate interplay of multiple modalities and the abstract nature of the three stimuli types. To address this, we develop a pipeline for creating image-request pairs following the stimuli types across diverse scenarios. This pipeline employs a two-step generation process: <b>candidate generation</b> and <b>candidate filtering</b>.
<ul>
<li><b>candidate generation</b>: we leveraged LLMs to generate diverse scenarios by providing the LLMs with a carefully crafted prompt, incorporating several oversensitivity samples as exemplars.</li>
<li><b>candidate filtering</b>: reviewers from Amazon Mechanical Turk evaluate the harmfulness and naturalness of our samples.</li>
</ul>
You can download the dataset on <a href="https://huggingface.co/datasets/AIcell/MOSSBench" target="_blank">Hugging Face Dataset</a>.
</p>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column" style="margin-right: -20rem;">
<div class="content has-text-centered">
<img src="website/static/images/statistic.png" alt="data-overview" style="max-width: 40%;"/>
<p>
Key statistics of <img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span>.<br/>
</p>
</div>
</div>
<div class="column">
<div class="content has-text-centered">
<img src="website/static/images/harm_type.png" alt="data-composition" style="max-width: 30%;"/>
<p>
Distribution of <img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span>.<br/>
</p>
</div>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-max-desktop has-text-centered">
<h2 class="title is-3" id="visualization">Visualization</h2>
<iframe src="website/visualizer/explore.html" style="width: 100%;min-height: 50vh; border-radius: 20px;"></iframe>
</div>
</div>
</div>
</section>
<!-- RESULTS SECTION -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title is-1 mathvista">Experiment Results</h1>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<h2 class="title is-3">Results on SOTA MLLMs</h2>
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/main_results.png" alt="grade-lv" width="90%"/>
<p><b>Refusal rate</b> (% ↓) of MLLMs on <img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span> by GPT evaluation and human evaluation. All models are evaluated in deterministic zero-shot settings. The highest and lowest refusal rate among models in each section are highlighted in red and blue, respectively.
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/over_vs_contrast.png" alt="grade-lv" width="100%"/>
<p>
<b>Oversensitivity level versus Safety level</b> of MLLMs. The levels are decided by their refusal rate of samples. The higher models refuse harmful samples, the higher their safety levels are. The open-source models are marked in red, while the proprietary models are marked in blue.
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/system_prompt.png" alt="grade-lv" width="100%"/>
<p><b>Refusal rate</b> (% ↓) of Gemini-pro 1.5 with different system prompts. Adding different
instructions on default empty system prompt (vanilla), the other system prompts are incorporated
with different focus (helpfulness, scrutiny, and safety).
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/stage.png" alt="contexts" width="100%"/>
<p><b>Refusal rate</b> of caption types from Claude 3-opus and Gemini-pro 1.5 on 30 refusal-triggering instances across different stimulus types: Exaggerated Risk (ER), Negated Harm (NH), and Counterintuitive Interpretation (CI). The responses are categorized into Correct Perception, Safety-critical Misperception, and Refusal.
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/oracle.png" alt="geometric reasoning" width="100%"/>
<p><b>Refusal rate</b> (% ↓) of Intent Reasoning and Safety Judgement evaluation on Claude 3
opus and Gemini-pro 1.5. Average refusal rate is shown in red dashed line, while Exaggerated risk,
Negated Harm, and Counterintuitive Interpretation are in Purple, Green, and Blue.</p>
</div>
</div>
</div>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<h2 class="title is-3">Oversensitivity and Compliance Examples</h2>
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/Exaggerated.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/Negated.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/Counterintuitive.png" alt="" width="60%"/>
</div>
</div>
</div>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<h2 class="title is-3">Cognitive Distortion in MLLMs Examples</h2>
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/moralizing.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/moralizing_2.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/moralizing_3.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/Catastrophizing.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/Catastrophizing_2.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/mental_filtering.png" alt="" width="60%"/>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="website/static/images/results-examples/mental_filtering_2.png" alt="" width="60%"/>
</div>
</div>
</div>
</div>
</div>
<div class="container is-full has-text-centered content m-6" id="result-table">
<h2 class="title is-3" id="explorer">Explorer</h2>
<p>Explore the outputs of each model on <img src="website/static/images/psych.webp" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">MOSSBench</span></p>
<div class="level has-text-centered" style="position: sticky; top: 0; z-index: 20;">
<div class="level-item box m-3" style="width: 30%; background: rgba(250, 250, 250, 1);">
<button class="button" style="width: 100%; border: none; background: rgba(250, 250, 250, 1);" id="refresh-qids">
<span class="icon is-large">
<i class="fa fa-redo fa-lg" aria-hidden="true"></i>
</span>
<p class="title is-4 m-0">Refresh Question</p>
</button>
</div>
<div class="level-item box m-3" style="width: 30%; background: rgba(250, 250, 250, 1);">
<div class="dropdown" style="width: 100%;">
<div class="dropdown-trigger has-text-justified" style="width: 100%; ">
<button class="button" aria-haspopup="true" aria-controls="dropdown-menu" style="width: 100%; border: none; background: rgba(250, 250, 250, 1);">
<p class="title m-0 is-4 dropdown-display">Gemini-Pro 1.5</p>
<span class="icon is-large" style="position: absolute; right:0;">
<i class="fas fa-angle-down fa-lg" aria-hidden="true"></i>
</span>
</button>
</div>
<div class="dropdown-menu" id="dropdown-menu" role="menu" style="width:100%;">
<div class="dropdown-content">
<!-- <a class="dropdown-item">
Dropdown item
</a>
<a class="dropdown-item">
Other dropdown item
</a> -->
</div>
</div>
</div>
</div>
<div class="level-item box m-3" style="width: 30%; background: rgba(250, 250, 250, 1);">
<div class="dropdown" style="width: 100%;">
<div class="dropdown-trigger has-text-justified" style="width: 100%;">
<button class="button" aria-haspopup="true" aria-controls="dropdown-menu" style="width: 100%; border: none; background: rgba(250, 250, 250, 1);">
<p class="title m-0 is-4 dropdown-display">Claude 3 Opus</p>
<span class="icon is-large" style="position: absolute; right:0;">
<i class="fas fa-angle-down fa-lg" aria-hidden="true"></i>
</span>
</button>
</div>
<div class="dropdown-menu" id="dropdown-menu" role="menu" style="width:100%;">
<div class="dropdown-content">
<!-- <a class="dropdown-item">
Dropdown item
</a>
<a class="dropdown-item">
Other dropdown item
</a> -->
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- @PAN TODO: bibtex -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title is-3 has-text-centered">BibTeX</h2>
<pre><code>@misc{li2024mossbenchmultimodallanguagemodel,
title={MOSSBench: Is Your Multimodal Language Model Oversensitive to Safe Queries?},
author={Xirui Li and Hengguang Zhou and Ruochen Wang and Tianyi Zhou and Minhao Cheng and Cho-Jui Hsieh},
year={2024},
eprint={2406.17806},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2406.17806},
}</code></pre>
</div>
</section>
<section>
<div class="section" id="org-banners" style="display:flex">
<a href="https://www.ucla.edu/" target="blank" rel="external">
<img class="center-block org-banner" src="website/static/images/ucla.png">
</a>
<a href="https://umd.edu/" target="blank" class="ext-link">
<img class="center-block org-banner" src="website/static/images/UMD.png">
</a>
<a href="https://www.psu.edu/" target="blank" rel="external">
<img class="center-block org-banner" src="website/static/images/pennstate.png">
</a>
</div>
</section>
<footer class="footer">
<!-- <div class="container"> -->
<div class="content has-text-centered">
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is website adapted from <a href="https://nerfies.github.io/">Nerfies</a> and <a href="https://mathvista.github.io/">MathVista</a>.
</p>
<p>Licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</div>
<!-- </div> -->
</footer>
</body>
</html>