-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
695 lines (639 loc) · 33.6 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Evaluating mathematical reasoning of foundation models in visual contexts">
<meta name="keywords" content="MathVista, Math Vista">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title> PhysBench</title>
<link rel="icon" href="./static/images/GVL.png">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="stylesheet" href="./static/css/leaderboard.css">
<script type="text/javascript" src="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="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/explorer-index.js"></script>
<script src="./static/js/question_card.js"></script>
<!-- Table-->
<script src="./static/js/leaderboard_physbench_test.js"></script>
<!-- Table data-->
<script src="./data/physbench_test/model_scores.js" defer></script>
<script src="./data/physbench_test/model_scores_nonv.js" defer></script>
</head>
<body>
<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="static/images/physbench.png" style="width:1em;vertical-align: middle" alt="Logo"/>
<span class="mathvista" style="vertical-align: middle">PhysBench</span>
</h1>
<h2 class="subtitle is-3 publication-subtitle">
Benchmarking and Enhancing VLMs for <br> Physical World Understanding
</h2>
<div class="is-size-5 publication-authors">
<!-- <span class="author-block">Anonymous</span> -->
<span class="author-block">
<a href="None">Wei Chow</a><sup style="color:#6fbf73;">1</sup><sup>*</sup>,</span>
<span class="author-block">
<a href="https://pointscoder.github.io/">Jiageng Mao</a><sup style="color:#6fbf73;">1</sup><sup>*</sup>,</span>
<span class="author-block">
<a href="https://sites.google.com/site/boyilics/home">Boyi Li</a><sup style="color:#ffac33;">2</sup>,
</span>
<span class="author-block">
<a href="https://danielseita.github.io/">Daniel Seita</a><sup style="color:#6fbf73">1</sup>,
</span>
<span class="author-block">
<a href="https://scholar.google.com.br/citations?user=ow3r9ogAAAAJ&hl=en">Vitor Guizilini</a><sup style="color:#ac33ff">3</sup>,
</span>
<span class="author-block">
<a href="https://yuewang.xyz/">Yue Wang</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 Southern California,</span>
<span class="author-block"><sup style="color:#ffac33">2</sup>UC Berkeley,</span
<span class="author-block"><sup style="color:#ac33ff">3</sup>Toyota Research Institute</span><br>
<span class="author-block"><sup>*</sup>Equal Contribution</span><br>
<!-- <span class="paper-block"><b style="color:#f41c1c">under review for ICLR 2025</b> </span> -->
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="TODO"
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>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/USC-GVL/PhysBench"
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/USC-GVL/PhysBench"
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>
<!-- Leaderboard Link. -->
<span class="link-block">
<a href="https://physbench.github.io/#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>
<!-- Eval.AI Link -->
<span class="link-block">
<a href="https://eval.ai/web/challenges/challenge-page/2287/overview" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<p style="font-size:18px">🔺</p>
<!-- <img src="https://eval.ai/dist/images/evalai-logo-single.png" alt="EvalAI Logo" style="height: 18px;"> -->
</span>
<span>Eval.AI</span>
</a>
</span>
</div>
</div>
</div>
</div>
</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 id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/tease_scores.png" alt="geometric reasoning" width="84%"/>
<p>
(a) The performance of 8 representative open-source VLMs across 19 sub-tasks in <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>. The closer it is to the circular boundary, the better.
<br>
(b) The overall performance of those 8 VLMs. Closed-source models generally perform better.
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/teaser.png" alt="geometric reasoning" width="84%"/>
<p>
<b>Common VQA</b> tasks typically involve questions about visual content and general knowledge.
<br>
<img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> emphasizes understanding the physical world, encompassing 4 dimensions.
</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>
<b>Vision-Language Models (VLMs)</b> have emerged as promising tools for building <b>embodied agents</b>, whereas their lack of <b>physical world understanding</b> hampers their effectiveness in real-world applications. To address this challenge, we present <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>, a comprehensive benchmark designed to evaluate and enhance VLMs' understanding of the physical world across diverse and complex tasks.
</p>
<p>
PhysBench comprises <b>100,000 entries</b> of interleaved video-image-text data, and the data is categorized into four major classes: <b>physical object properties</b>, <b>physical object relationships</b>, <b>physical scene understanding</b>, and <b>physics-driven dynamics</b>, covering <b>19 subclasses</b> and <b>10 distinct capability dimensions</b>.
</p>
<p>
Our extensive experiments on 39 representative VLMs reveal significant gaps in physical world understanding, likely due to the absence of physical knowledge in their training data.
To improve VLMs' physical understanding, we propose an agent-based method called <b>PhysAgent</b>, which leverages prior physical knowledge and expert model assistance to enhance physical world understanding capabilities.
</p>
<p>
Furthermore, we demonstrate that improving VLMs’ understanding of the physical world can significantly facilitate the deployment of embodied agents in real-world scenarios, moving towards bridging the gap between human and machine intelligence in comprehending the physical world.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered">
<div class="column is-full has-text-centered content">
<h2 class="title is-3" id="leaderboard">Leaderboard on PhysBench</h2>
<div class="content">
<p class="mt-3">Accuracy scores for <strong>General VLM</strong> on the <b>test</b> subset (10,002 entries) of <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>.
</p>
<div id="physbench_test"></div>
</div>
<div class="content">
<p class="mt-3">Accuracy scores for <strong>Image VLM</strong> and <strong>Video VLM</strong> on the <b>test</b> subset without interleaved data entries (8,099 entries) of <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>.
</p>
<div id="physbench_test_nonv"></div>
</div>
<div class="content">
<br>
<b>Method types:</b> <b>Seq ⏩:</b> Sequential input of images after frame selection from videos,, <b>Merge 🖼️:</b> merging
video frames into a single image
<br>
<br>
<div>
<p>🚨 To submit your results to the leaderboard, please send to <a href="mailto:zhouwei_haining@outlook.com">this email</a> with your result json files.</p>
<p>🚨 For more submission details, please refer to <a href="https://github.com/lupantech/MathVista?tab=readme-ov-file#-leaderboard-">this link</a> and <a href="https://github.com/lupantech/MathVista?tab=readme-ov-file#-evaluations-on-mathvista">this link</a>.
</p>
</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="static/images/physbench.png" style="width:1em;vertical-align: middle" alt="Logo"/>
<span class="mathvista" style="vertical-align: middle">PhysBench 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>
We propose <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>, which comprehensively evaluates VLMs' perception of the physical world across four major task categories:
</p>
<ul>
<li>
<b>(1) Physical Object Properties</b>: Assessment of physical attributes of objects such as mass, size, density, tension, friction, bending stiffness, elasticity, and plasticity.
</li>
<li>
<b>(2) Physical Object Relationships</b>: Evaluation of spatial relationships involving object movement, speed, and position.
</li>
<li>
<b>(3) Physical Scene Understanding</b>: Interpretation of environmental factors, including light sources, viewpoints, temperature, and so on.
</li>
<li>
<b>(4) Physics-based Dynamics</b>: Understanding of physical events like collisions, throwing, fluid dynamics, explosions, and similar phenomena.
</li>
</ul>
<div class="content has-text-centered">
<img src="static/images/part1/data_cases_full.png" alt="algebraic reasoning" width="80%"/>
<p>Sampled PhysBench examples from four major dimensions</p>
<br>
</div>
<div class="content has-text-centered">
<img src="static/images/part1/related.png" alt="arithmetic reasoning" width="80%"/>
<br>
A comparison between <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>. and other physical understanding question-answering benchmarks reveals that
<br><img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> is a comprehensive dataset, covering a wide range of tasks related to physical world understanding.
</div>
<p>
The complete <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> dataset consists of 100,000 entries, organized into 19 subclasses and 10 distinct capability dimensions.
For convenience, we selected a subset of 10,002 entries, which are more challenging and diverse, as the test set, and 200 entries as the validation set for parameter choosing.
<ul>
<li><b>val</b>: 200 examples used for model development, validation, or for those with limited computing resources.</li>
<li><b>test</b>: 10,002 examples for standard evaluation (include val). Notably,
the answer labels for test will NOT be publicly released.</li>
<li><b>train</b>: The remaining 89,998 examples.</li>
</ul>
You can download the dataset on <a href="https://huggingface.co/datasets/AI4Math/MathVista" target="_blank">Hugging Face Dataset</a>.
</p>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column">
<div class="content has-text-centered">
<img src="static/images/part1/stat.png" alt="data-overview" style="max-width: 75%;"/>
<p>
Key statistics of <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>.<br/>
</p>
</div>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<h2 class="title is-3">Statistics</h2>
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/part1/stat1.png" alt="qs-len" class="stats-image"/>
<p>
The distribution of the number of words per question in <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>.
<br> Questions with a length greater than 48 are categorized as 47 for visualization simplicity.
<!-- Questions with a length greater than 60 are categorized as 61 for visualization simplicity -->
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/part1/stat2.png" alt="reasoning" class="stats-image"/>
<p>The distribution of the number of words per question in <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>.
<br>
Options with a length greater than 20 are categorized as 20 for visualization simplicity.
</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/part1/stat3.png" alt="reasoning-count" class="stats-image"/>
<p>The distribution charts for image and video resolution, as well as video frame counts.
<br>From left to right: the distribution of image resolution, <br>the distribution of video resolution, and the distribution of video frame counts.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- RESULTS SECTION -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h3 class="title is-3 mathvista">Can VLMs Understand the Physical World?</h3>
</div>
</section>
<section class="section">
<div class="container">
<!-- For words!-->
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
To assess whether VLMs can understand the physical world, we evaluated 39 representative VLMs on <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> and found that:
</p>
<ul>
<li><b>VLMs exhibit limited understanding of the physical world</b>.</li>
<li><b>Closed-source models generally perform better</b>.</li>
</ul>
</div>
</div>
<br>
<div class="content has-text-centered">
<img src="static/images/question1/cor.png" alt="grade-lv" width="80%"/>
<p>
The visualization of model performance across 19 sub-tasks is presented, where different colors represent the respective categories.
<br>
The four colors, from left to right, represent physical object properties, physical object relationships, physical scenes, and physical-based dynamics.
</p>
</div>
</div>
</section>
<!-- RESULTS SECTION -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h3 class="title is-3 mathvista">Why Do VLMs Struggle with Physical World Understanding?</h3>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
To assess whether VLMs can understand the physical world, we evaluated 39 representative VLMs on <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> and found that:
</p>
<ul>
<li><b>Physical world understanding differs significantly from common VQA tasks</b>. We established a correlation map with the common VQA benchmark. Our analysis identifies a notable distinction between PhysBench and traditional VLM benchmarks, with PhysBench demonstrating a closer alignment with MMMU, which necessitates complex reasoning and diverging from the majority of other benchmarks.</li>
<li><b>VLMs's physical world understanding ability does not scale with model size, data, or frames</b>.While keeping the data size constant, increasing the model size or, alternatively, increasing the data size while keeping the model size unchanged, led to inconsistent results. Similarly, increasing the number of frames also yielded unstable outcomes.</li>
</ul>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/cor.png" alt="grade-lv" width="60%"/>
<p>Correlation map between 4 tasks in <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>. and 15 other vision-language benchmarks.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/model.png" alt="grade-lv" width="60%"/>
<p>Model Size Scalability.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/data.png" alt="grade-lv" width="60%"/>
<p>Data Scalability.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/frame.png" alt="grade-lv" width="60%"/>
<p>Frame Scalability.</p>
</div>
</div>
</div>
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
We were perplexed by the fact that increasing the amount of training data did not improve the VLM's understanding of the physical world.
To investigate further, we examined the training datasets of LLaVA-1.5, VILA-1.5, and PLLaVA-1.5 and identified a lack of physical world knowledge in these datasets.
Additionally, keywords frequently encountered in <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> are notably rare in the training data of these model.
<b>This deficiency in relevant data likely contributes to the VLM's poor comprehension of physical world concepts</b>.
We further support this hypothesis by analyzing the error distribution and fine-tuning the VLM in subsequent experiments.
</p>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/word1.png" alt="grade-lv" width="80%"/>
<p>Word Statics and Word Cloud for <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/word2.png" alt="grade-lv" width="80%"/>
<p>Word Statics and Word Cloud for LLaVA-1.5-13B Training Data.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/word3.png" alt="grade-lv" width="80%"/>
<p>Word Statics and Word Cloud for VILA-1.5-13B Training Data.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/word4.png" alt="grade-lv" width="80%"/>
<p>Word Statics and Word Cloud for PLLaVA-13B Training Data.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question2/word5.png" alt="grade-lv" width="80%"/>
<p>The frequency of common terms in <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span> within the training data of the LLaVA-1.5-13B, VILA-1.5-13B, and PLLaVA-13B models.</p>
</div>
</div>
</div>
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
To investigate the poor performance of VLMs on <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>, we randomly selected 500 questions and obtained explanations from
three models—GPT-4o, Phi-3V, and Gemini-1.5-flash. Expert annotators classified the root causes
of the mispredictions into six categories: perception errors, reasoning errors, lack of knowledge,
refusal to answer, failure to follow instructions, and annotation errors in the dataset.
We find that <b>perceptual and knowledge gaps constitute the majority of errors</b>.
</p>
</div>
</div>
<br>
<div class="content has-text-centered">
<img src="static/images/question2/error.png" alt="grade-lv" width="70%"/>
<p>Distribution of error types for GPT-4V, Gemini-1.5-flash, Phi-3V.</p>
</div>
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
Our error analysis revealed that inadequate physical
world knowledge and reasoning capabilities were key contributors to the models’ poor performance.
To investigate whether introducing additional examples could enhance performance, we
conducted tests on 200 entries of <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>, pairing each with a similar example. These additional
examples were incorporated through fine-tuning or in-context learning. As shown in the below figure,
the performance improvements after adding physical world knowledge examples indicate that VLMs
can transfer physical knowledge to some extent. This suggests that the original data’s lack of physical
world knowledge was a significant factor in the models’ suboptimal performance.
</p>
</div>
</div>
<div class="content has-text-centered">
<img src="static/images/question2/transfer.png" alt="grade-lv" width="50%"/>
<p>Physics knowledge transfer study.</p>
</div>
</div>
</section>
<!-- RESULTS SECTION -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h3 class="title is-3 mathvista">How to enhance VLMs for Physical World Understanding?</h3>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
Recognizing perceptual inaccuracies and knowledge gaps as key sources of error, we introduce <b>PhysAgent</b> to improve VLMs' understanding of the physical world by integrating expert models for enhanced perception and incorporating memory for physical knowledge.
</p>
</div>
</div>
<br>
<div class="content has-text-centered">
<img src="static/images/question3/physagent.png" alt="grade-lv" width="80%"/>
<p>Architecture of PhysAgent.</p>
</div>
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
The results lead to the following conclusions:
<br>
<b>(1) Prompting methods is unstable, and using pure language yields catastrophic results</b>.
As observed, the CoT strategy has minimal impact, while both Desp-CoT and PLR show a decline in performance.
This suggests that descriptive prompts are not particularly effective for addressing the questions,
implying that our dataset requires a deeper understanding of the videos or images to answer accurately.
<br>
<b>(2) ContPhy even worsens performance</b>.
In three out of four tasks, ContPhy underperforms compared to its base model,
GPT-4o, due to suboptimal module invocation and limited flexibility in its logical templates,
which struggle to adapt to diverse scenarios. Additionally, ContPhy relies on models
like RCNN to process visual information instead of directly leveraging GPT-4o,
leading to potential information loss and subsequent performance degradation.
<br>
<b>(3) PhysAgent consistently improves zero-shot performance</b>, notably achieving a 36.5% improvement
for GPT-4o in spatial tasks. Compared to the CoT, Desp-CoT, and PLR prompting strategies,
our method demonstrates clear advantages.
</p>
</div>
</div>
<br>
<div class="content has-text-centered">
<img src="static/images/question3/exp.png" alt="grade-lv" width="30%"/>
<p>Performance of different methods.</p>
</div>
</div>
</section>
<!-- RESULTS SECTION -->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h3 class="title is-3 mathvista">Can Physical World Understanding Help Embodied Application?</h3>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
we conducted five embodied agent tasks to verify <b>enhancing VLMs' physical understanding facilitates the deployment of embodied agents</b>.
</p>
</div>
</div>
<div class="columns is-centered m-6">
<div class="column is-full has-text-centered content">
<div id="results-carousel" class="carousel results-carousel">
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question4/task.png" alt="grade-lv" width="60%"/>
<p>Description of each of the testing tasks.</p>
</div>
</div>
<div class="box m-5">
<div class="content has-text-centered">
<img src="static/images/question4/moka_s.png" alt="grade-lv" width="80%"/>
<p>Marked observation, predicted affordances
and motion in MOKA. MOKA leverages a VLM to generate motions based on a point-based
affordance representation.</p>
</div>
</div>
</div>
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="content has-text-justified">
<p>
we observe consistent improvements after fine-tuning with a subset of <img src="static/images/physbench.png" style="width:1.0em;vertical-align: middle" alt="Logo"/>
<span class="mathvista">PhysBench</span>,
indicating that the benchmark's data is of high quality and suitable for use as demonstration data in
open-world robotics tasks. Additionally,
PhysAgent consistently yields stable zero-shot gains across all five tasks,
with especially significant progress observed in the force task.
While the improvements are less pronounced compared to direct fine-tuning.
</p>
</div>
</div>
<br>
<div class="content has-text-centered">
<img src="static/images/question4/exp.png" alt="grade-lv" width="40%"/>
<p>Performance on 5 embodied tasks shown in the former figure.
<br>The color blocks from left to right represent success, VLM reasoning error, and execution error, respectively.</p>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title is-3 has-text-centered">BibTeX</h2>
<pre><code>
@inproceedings{
TODO
}
</code></pre>
</div>
</section>
<footer class="footer">
<div class="content has-text-centered">
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p style="font-size: 14px;">
This website is adapted from <a href="https://nerfies.github.io/">Nerfies</a> and <a href="https://mathvista.github.io/">MathVista</a>, 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>
</footer>
</body>
</html>