-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
666 lines (605 loc) · 34.4 KB
/
index.html
File metadata and controls
666 lines (605 loc) · 34.4 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<!-- Primary Meta Tags -->
<meta name="title" content="Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders - ICLR 2026">
<meta name="description" content="We show encoder redundancy in multi-encoder MLLMs: performance degrades gracefully when encoders are masked. We introduce CUR and IG metrics and show dual-encoder variants achieve >90% performance at lower cost.">
<meta name="keywords" content="multimodal large language models, vision encoders, encoder redundancy, MLLM, CUR, information gap, computer vision, AI">
<meta name="author" content="Yizhou Wang, Song Mao, Yang Chen, Yufan Shen, Yinqiao Yan, Pinlong Cai, Ding Wang, Guohang Yan, Zhi Yu, Xuming Hu, Botian Shi">
<meta name="robots" content="index, follow">
<meta name="language" content="English">
<!-- Open Graph / Facebook -->
<meta property="og:type" content="article">
<meta property="og:site_name" content="Shanghai Artificial Intelligence Laboratory">
<meta property="og:title" content="Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders">
<meta property="og:description" content="We show encoder redundancy in multi-encoder MLLMs: performance degrades gracefully when encoders are masked. We introduce CUR and IG metrics and show dual-encoder variants achieve >90% performance at lower cost.">
<meta property="og:url" content="https://encoder-redundancy.github.io/">
<meta property="og:image" content="https://encoder-redundancy.github.io/static/images/all_models_scores.png">
<meta property="og:image:width" content="1200">
<meta property="og:image:height" content="630">
<meta property="og:image:alt" content="Encoder Redundancy in MLLMs - Research Preview">
<meta property="article:published_time" content="2026-01-01T00:00:00.000Z">
<meta property="article:author" content="Yizhou Wang">
<meta property="article:section" content="Research">
<meta property="article:tag" content="MLLM">
<meta property="article:tag" content="encoder redundancy">
<!-- Twitter -->
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:title" content="Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders">
<meta name="twitter:description" content="We show encoder redundancy in multi-encoder MLLMs: performance degrades gracefully when encoders are masked. We introduce CUR and IG metrics.">
<meta name="twitter:image" content="https://encoder-redundancy.github.io/static/images/all_models_scores.png">
<meta name="twitter:image:alt" content="Encoder Redundancy in MLLMs - Research Preview">
<!-- Academic/Research Specific -->
<meta name="citation_title" content="Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders">
<meta name="citation_author" content="Wang, Yizhou">
<meta name="citation_author" content="Mao, Song">
<meta name="citation_author" content="Chen, Yang">
<meta name="citation_publication_date" content="2026">
<meta name="citation_conference_title" content="International Conference on Learning Representations (ICLR)">
<meta name="citation_pdf_url" content="https://encoder-redundancy.github.io/static/pdfs/paper.pdf">
<!-- Additional SEO -->
<meta name="theme-color" content="#2563eb">
<meta name="msapplication-TileColor" content="#2563eb">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="default">
<!-- Preconnect for performance -->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link rel="preconnect" href="https://ajax.googleapis.com">
<link rel="preconnect" href="https://documentcloud.adobe.com">
<link rel="preconnect" href="https://cdn.jsdelivr.net">
<title>Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders | ICLR 2026</title>
<!-- Critical CSS - Load synchronously -->
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/index.css">
<!-- Non-critical CSS - Load asynchronously -->
<link rel="preload" href="static/css/bulma-carousel.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="static/css/bulma-slider.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="static/css/fontawesome.all.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<!-- Fallback for browsers that don't support preload -->
<noscript>
<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">
</noscript>
<!-- Fonts - Optimized loading -->
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap" rel="stylesheet">
<!-- MathJax 3 for equation rendering (LaTeX: $...$ inline, $$...$$ display) -->
<script>
window.MathJax = {
tex: {
inlineMath: [['$', '$'], ['\\(', '\\)']],
displayMath: [['$$', '$$'], ['\\[', '\\]']],
processEscapes: true
},
options: {
skipHtmlTags: ['script', 'noscript', 'style', 'textarea', 'pre', 'code']
}
};
</script>
<script src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js" async></script>
<!-- Defer non-critical JavaScript -->
<script defer src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script defer src="static/js/bulma-carousel.min.js"></script>
<script defer src="static/js/bulma-slider.min.js"></script>
<script defer src="static/js/index.js"></script>
<!-- Structured Data for Academic Papers -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "ScholarlyArticle",
"headline": "Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders",
"description": "We show encoder redundancy in multi-encoder MLLMs and introduce CUR and IG metrics. Dual-encoder variants achieve over 90% of full performance at significantly reduced cost.",
"author": [
{"@type": "Person", "name": "Yizhou Wang", "affiliation": {"@type": "Organization", "name": "Shanghai AI Lab"}},
{"@type": "Person", "name": "Song Mao", "affiliation": {"@type": "Organization", "name": "Shanghai AI Lab"}},
{"@type": "Person", "name": "Yang Chen", "affiliation": {"@type": "Organization", "name": "Shanghai AI Lab"}},
{"@type": "Person", "name": "Yufan Shen", "affiliation": {"@type": "Organization", "name": "Shanghai AI Lab"}},
{"@type": "Person", "name": "Botian Shi", "affiliation": {"@type": "Organization", "name": "Shanghai AI Lab"}}
],
"datePublished": "2026",
"publisher": {"@type": "Organization", "name": "ICLR"},
"url": "https://encoder-redundancy.github.io/",
"image": "https://encoder-redundancy.github.io/static/images/all_models_scores.png",
"keywords": ["MLLM", "vision encoders", "encoder redundancy", "CUR", "information gap", "multimodal"],
"abstract": "Recent multimodal large language models (MLLMs) increasingly integrate multiple vision encoders to improve performance. We show this assumption often fails: performance typically degrades gracefully—and sometimes improves—when selected encoders are masked, revealing pervasive encoder redundancy. We introduce Conditional Utilization Rate (CUR) and Information Gap (IG) to quantify this effect. Single- and dual-encoder variants recover over 90% of baseline on most non-OCR tasks with substantially lower training and inference cost.",
"isAccessibleForFree": true,
"license": "https://creativecommons.org/licenses/by/4.0/",
"about": [{"@type": "Thing", "name": "Multimodal LLMs"}, {"@type": "Thing", "name": "Vision Encoders"}]
}
</script>
<!-- Website/Organization Structured Data -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Shanghai Artificial Intelligence Laboratory",
"url": "https://www.shlab.org.cn",
"sameAs": ["https://github.com/Encoder-Redundancy"]
}
</script>
</head>
<body>
<!-- Scroll to Top Button -->
<button class="scroll-to-top" onclick="scrollToTop()" title="Scroll to top" aria-label="Scroll to top">
<i class="fas fa-chevron-up"></i>
</button>
<!-- More Works Dropdown -->
<div class="more-works-container">
<button class="more-works-btn" onclick="toggleMoreWorks()" title="View More Works from Our Lab">
<i class="fas fa-flask"></i>
More Works
<i class="fas fa-chevron-down dropdown-arrow"></i>
</button>
<div class="more-works-dropdown" id="moreWorksDropdown">
<div class="dropdown-header">
<h4>More Works from Our Lab</h4>
<button class="close-btn" onclick="toggleMoreWorks()">
<i class="fas fa-times"></i>
</button>
</div>
<div class="works-list">
<a href="https://alpha-innovator.github.io/DocGenome_page/" class="work-item" target="_blank">
<div class="work-info">
<h5>DocGenome: An Open Large-scale Scientific Document Benchmark for Training and Testing Multi-modal Large Language Models</h5>
<p>A large-scale benchmark for training and testing multi-modal LLMs.</p>
</div>
<i class="fas fa-external-link-alt"></i>
</a>
</div>
</div>
</div>
<main id="main-content">
<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">Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders</h1>
<div class="is-size-5 publication-authors">
<span class="author-block"><a href="https://github.com/1zhou-Wang" target="_blank">Yizhou Wang</a><sup>*,†</sup>,</span>
<span class="author-block"><a href="https://maosong.website/" target="_blank">Song Mao</a><sup>*</sup>,</span>
<span class="author-block">Yang Chen<sup>*,†</sup>,</span>
<span class="author-block">Yufan Shen,</span>
<span class="author-block">Yinqiao Yan,</span>
<span class="author-block">Pinlong Cai,</span>
<span class="author-block">Ding Wang,</span>
<span class="author-block">Guohang Yan,</span>
<span class="author-block">Zhi Yu,</span>
<span class="author-block">Xuming Hu,</span>
<span class="author-block">Botian Shi</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">Shanghai AI Lab, HKUST(GZ), Zhejiang University, Beijing University of Technology<br>ICLR 2026</span>
<span class="eql-cntrb"><small><br><sup>*</sup>Equal Contribution. † Work done during internship at Shanghai AI Lab.</small></span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/abs/2507.03262" target="_blank" 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>
<span class="link-block">
<a href="https://github.com/1zhou-Wang/Investigating_MultiEncoder_Redundancy" target="_blank" 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>
<span class="link-block">
<a href="https://openreview.net/forum?id=cAopJVLKvi" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="ai ai-openreview"></i></span>
<span>OpenReview</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/collections/DiDisama/finetuned-eagle-models" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fab fa-hugging-face"></i></span>
<span>Models</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Teaser figure -->
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<img src="static/images/Redundancy_illustration.jpg" alt="Illustration of encoder redundancy in multi-encoder MLLMs" style="max-width:100%; height:auto;" loading="eager"/>
<h2 class="subtitle has-text-centered">
Different vision encoders provide similar or conflicting visual cues; ablating one or several encoders can maintain or even improve performance, revealing pervasive encoder redundancy in multi-encoder MLLMs.
</h2>
</div>
</div>
</section>
<!-- End teaser -->
<!-- Paper abstract -->
<!-- <section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Recent multimodal large language models (MLLMs) increasingly integrate multiple vision encoders to improve performance on various benchmarks, assuming that diverse pretraining objectives yield complementary visual signals. However, we show this assumption often fails in practice. Through systematic encoder masking across representative multi-encoder MLLMs, we find that performance typically degrades gracefully—and sometimes even improves—when selected encoders are masked, revealing pervasive encoder redundancy.
</p>
<p>
To quantify this effect, we introduce two principled metrics: the <strong>Conditional Utilization Rate (CUR)</strong>, which measures an encoder's marginal contribution in the presence of others, and the <strong>Information Gap (IG)</strong>, which captures heterogeneity in encoder utility within a model. Using these tools, we observe: (i) strong specialization on tasks like OCR & Chart, where a single encoder can dominate with a CUR >90%; (ii) high redundancy on general VQA and knowledge-based tasks, where encoders are largely interchangeable; (iii) instances of detrimental encoders with negative CUR. Notably, masking specific encoders can yield up to 16% higher accuracy on a specific task category and 3.6% overall performance boost compared to the full model.
</p>
<p>
Furthermore, single- and dual-encoder variants recover over 90% of baseline on most non-OCR tasks with substantially lower training resources and inference latency. Our analysis challenges the "more encoders are better" heuristic in MLLMs and provides actionable diagnostics for developing more efficient and effective multimodal architectures.
</p>
</div>
</div>
</div>
</div>
</section> -->
<!-- End paper abstract -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Motivation</h2>
<div class="content has-text-justified">
<p>
There are two primary strategies to enhance the perception capabilities of MLLMs:
</p>
<ul>
<li><strong>Improve a single encoder</strong>: Increase parameters, fuse features across layers, etc.</li>
<li><strong>Integrate multiple encoders</strong>: Combine encoders to potentially capture complementary visual signals.</li>
</ul>
<p>
This work investigates the second approach—exploring the efficiency and effectiveness of multi-encoder MLLMs.
</p>
<p>
Notably, the aggregate performance of multi-encoder MLLMs is often not simply the sum of individual encoder capabilities.
</p>
</div>
<!-- Case Study -->
<h3 class="title is-4 mt-5">Case Study</h3>
<div class="content has-text-justified">
<p>
Masking a task-critical encoder can dramatically alter model outputs—demonstrating that multi-encoder MLLMs may depend heavily on specific encoders for certain tasks. Below, we show examples using Eagle-X4 8B Plus with and without selected encoders:
</p>
</div>
<div class="content has-text-centered" style="max-width:900px; margin: 0 auto;">
<img src="static/images/long_case_1.jpg" alt="Case study 1: Eagle-X4 8B with/without specific encoders masked" loading="lazy" style="width:100%; max-width:900px; height:auto; margin-bottom: 0.75rem;"/>
<p class="is-size-7 mt-3">Case study 1.</p>
<img src="static/images/long_case_2.jpg" alt="Case study 2: encoder masking effect on model output" loading="lazy" style="width:100%; max-width:900px; height:auto; margin-top: 1.75rem;"/>
<p class="is-size-7 mt-3">Case study 2.</p>
</div>
<div class="content has-text-justified">
<p>
As these examples show, some encoders are redundant—ablating them often maintains model performance.
</p>
<p>
This observation raises several questions: How robust are multi-encoder MLLMs to encoder masking? How gracefully does performance degrade, and how can we quantify redundancy in these architectures?
</p>
</div>
</div>
</section>
<section class="section hero is-light">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Method</h2>
<div class="content has-text-justified">
<p>
<strong>What architectures do we study?</strong><br>
We focus on the widely adopted “ViT-adapter-LLM” family of multi-encoder MLLMs. As illustrated below: given an input image $I$ and prompt $T$, a model with vision encoders $\mathcal{E}_n = \{E_1,\dots,E_n\}$ produces an output $Y$ as follows:
</p>
<div style="text-align:center;">
$$
Y = f_{\mathcal{E}_n}(I, T) = \mathrm{LLM}\big(\mathrm{proj}(\mathrm{fusion}(E_1(I),\ldots,E_n(I))),\ T\big)
$$
</div>
<p>
Here, $\mathrm{fusion}(\cdot)$ refers to feature integration strategies such as concatenation or attention, and $\mathrm{proj}(\cdot)$ aligns visual features to the LLM’s embedding space via an adapter.<br>
The studied models include
<a href="https://github.com/NVlabs/Eagle" target="_blank" rel="noopener">Eagle</a>,
<a href="https://github.com/cambrian-mllm/cambrian" target="_blank" rel="noopener">Cambrian-1</a>,
<a href="https://github.com/tsb0601/MMVP" target="_blank" rel="noopener">I-MoF</a>,
<a href="https://github.com/NVlabs/Eagle" target="_blank" rel="noopener">Eagle2</a>,
<a href="https://github.com/deepseek-ai/DeepSeek-VL" target="_blank" rel="noopener">DeepSeek-VL</a>.
</p>
<p>
<strong>How can we quantify encoder redundancy?</strong><br>
To rigorously assess how individual encoders contribute—and overlap—within a multi-encoder MLLM, we introduce two intuitive and complementary metrics:
the <strong>Conditional Utilization Rate (CUR)</strong> and the <strong>Information Gap (IG)</strong>.
</p>
<ul style="margin-bottom: 0.75rem;">
<li>
<strong>Conditional Utilization Rate (CUR):</strong><br>
CUR measures the unique value provided by each encoder. For encoder $E_i$, it is calculated as:
<div style="text-align:center; margin: 0.5em 0;">
$$
u(E_i) = \frac{\mathrm{acc}\big(f_{\mathcal{E}_n}\big) - \mathrm{acc}\big(f_{\mathcal{E}_n \setminus \{E_i\}}\big)}{\mathrm{acc}\big(f_{\mathcal{E}_n}\big)}
$$
</div>
Here, $f_{\mathcal{E}_n}$ denotes the model with all encoders, while $f_{\mathcal{E}_n \setminus \{E_i\}}$ is the model with $E_i$ masked. A high $u(E_i)$ means $E_i$ is essential and uniquely valuable; values near zero indicate redundancy; negative values signal that an encoder may even harm overall performance.
</li>
<li style="margin-top: 1em;">
<strong>Information Gap (IG):</strong><br>
IG reflects how evenly the encoders contribute to the model. For a set of encoders $\mathcal{E}_n$:
<div style="text-align:center; margin: 0.5em 0;">
$$
\Delta_{\mathrm{gap}}(\mathcal{E}_n) = \max_{i \in \{1,\dots,n\}} u(E_i)\ -\ \min_{j \in \{1,\dots,n\}} u(E_j)
$$
</div>
A small $\Delta_{\mathrm{gap}}$ indicates balanced contributions among encoders, while a large gap points to significant imbalance—where some encoders dominate the model’s performance.
</li>
</ul>
<br>
<p>
Together, CUR and IG provide a clear, principled framework for diagnosing redundancy versus specialization among encoders in multi-encoder architectures.
</p>
<p>
<strong>How is performance measured?</strong><br>
We rigorously evaluate these multi-encoder MLLMs using a suite of benchmarks from <a href="https://github.com/open-compass/VLMEvalKit" target="_blank" rel="noopener">VLMEvalKit</a>, a comprehensive toolkit covering a wide range of multimodal LLM tasks. This ensures robust, standardized comparison across models and settings.<br><br>
<strong>Benchmarks used in evaluation:</strong>
</p>
<div class="table-container" style="max-width: 850px; margin: 0 auto;">
<table class="table is-bordered is-striped is-narrow is-hoverable" style="font-size:0.98em;">
<thead>
<tr>
<th style="min-width:104px;">Category</th>
<th style="min-width:138px;">Benchmark</th>
<th style="min-width:82px;">Metric</th>
<th style="min-width:115px;">Remark</th>
</tr>
</thead>
<tbody>
<!-- General -->
<tr>
<td rowspan="4" style="vertical-align:middle;text-align:center;"><strong>General</strong></td>
<td>
<a href="https://cs.stanford.edu/people/dorarad/gqa/about.html" target="_blank" title="GQA">GQA</a><sup>[1]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://opencompass.org.cn/dataset/mmbench" target="_blank" title="MMBench">MMB</a><sup>[2]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://github.com/BradyFU/Awesome-MME" target="_blank" title="MME">MME</a><sup>[3]</sup>
</td>
<td>Score</td>
<td>Perception score / 20</td>
</tr>
<tr>
<td>
<a href="https://github.com/AILab-CVC/SEED" target="_blank" title="SEED-I">SEED-I</a><sup>[4]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<!-- Knowledge -->
<tr>
<td rowspan="4" style="vertical-align:middle;text-align:center;"><strong>Knowledge</strong></td>
<td>
<a href="https://allenai.org/data/ai2d" target="_blank" title="AI2D">AI2D</a><sup>[5]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://mathvista.github.io/" target="_blank" title="MathVista">MathVista</a><sup>[6]</sup>
</td>
<td>Score</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://github.com/FranxYao/SQA" target="_blank" title="SQA-I">SQA-I</a><sup>[7]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://mmmu-benchmark.github.io/" target="_blank" title="MMMU">MMMU</a><sup>[8]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<!-- OCR & Chart -->
<tr>
<td rowspan="4" style="vertical-align:middle;text-align:center;"><strong>OCR & Chart</strong></td>
<td>
<a href="https://rrc.cvc.uab.es/?ch=17" target="_blank" title="DocVQA">DocVQA</a><sup>[9]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://github.com/vis-nlp/ChartQA" target="_blank" title="ChartQA">ChartQA</a><sup>[10]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://github.com/open-compass/OCRBench" target="_blank" title="OCRBench">OCRBench</a><sup>[11]</sup>
</td>
<td>Score</td>
<td>Score / 10</td>
</tr>
<tr>
<td>
<a href="https://textvqa.org/" target="_blank" title="TextVQA">TextVQA</a><sup>[12]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<!-- Vision-Centric -->
<tr>
<td rowspan="3" style="vertical-align:middle;text-align:center;"><strong>Vision-Centric</strong></td>
<td>
<a href="https://arxiv.org/abs/2402.12995" target="_blank" title="CV-Bench">CV-Bench</a><sup>[13]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://github.com/cambrian-mllm/MMVP" target="_blank" title="MMVP">MMVP</a><sup>[14]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
<tr>
<td>
<a href="https://github.com/grok-ai/Real_World_QA" target="_blank" title="Real World QA">Real World QA</a><sup>[15]</sup>
</td>
<td>Accuracy</td>
<td></td>
</tr>
</tbody>
</table>
</div>
<br>
</div>
</div>
</section>
<!-- Key Results: three subsections (no carousel) -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Key Results</h2>
<!-- Subsection 1: All models suffer from redundancy -->
<h3 class="title is-4">1. Pervasive Encoder Redundancy</h3>
<p class="content has-text-justified">
Across representative multi-encoder MLLMs (Eagle, Cambrian-1, I-MoF, Eagle2, DeepSeek-VL), performance degrades gracefully—and sometimes improves—when encoders are masked, showing that additional encoders often yield diminishing returns. Below: (left) overall performance vs. number of masked encoders; (right) Conditional Utilization Rate (CUR) by benchmark category—strong specialization on OCR & Chart vs. high redundancy on General and Knowledge.
</p>
<div class="mb-5">
<img src="static/images/all_models_scores.png" alt="Performance of multi-encoder MLLMs with different number of masked vision encoders" loading="lazy" style="width:100%; height:auto; display:block; margin:0 auto;"/>
<p class="has-text-centered is-size-7 mt-3">Performance with different numbers of masked encoders (Max / Min / Mean over subsets).</p>
</div>
<div class="mb-5">
<img src="static/images/cur_results.png" alt="CUR across benchmark categories" loading="lazy" style="width:100%; height:auto; display:block; margin:0 auto;"/>
<p class="has-text-centered is-size-7 mt-3">CUR by category: higher CUR = stronger dependence on that encoder; negative CUR = detrimental.</p>
</div>
<!-- Subsection 2: Re-trained vs original — efficiency and performance -->
<h3 class="title is-4 mt-5">2. Efficiency and Performance: Re-trained vs. Original</h3>
<p class="content has-text-justified">
Re-training with fewer encoders preserves most accuracy while substantially reducing cost. For Eagle-X5 7B, a dual-encoder variant (Eagle-X2 7B) reaches <strong>94%</strong> of the full model’s performance while cutting total training time by <strong>34%</strong> (on 8× A100). At inference, masking three encoders reduces latency by <strong>19.5%</strong> with <4% performance drop. Vision-side FLOPs drop to 61.4% when keeping only two encoders; dual-encoder variants consistently recover >90% of baseline on most non-OCR tasks with lower training and inference cost.
</p>
<div class="content table-container">
<table class="table is-bordered is-striped is-narrow" style="max-width: 640px; margin: 0 auto;">
<thead>
<tr>
<th>Aspect</th>
<th>Re-trained dual-encoder (e.g. Eagle-X2 7B) vs. full model</th>
</tr>
</thead>
<tbody>
<tr><td>Performance</td><td>≥94% of full model</td></tr>
<tr><td>Training time</td><td>~34% reduction</td></tr>
<tr><td>Inference latency</td><td>~19.5% reduction (when masking 3 encoders)</td></tr>
<tr><td>Vision FLOPs</td><td>61.4% of full model</td></tr>
</tbody>
</table>
</div>
</section>
<!-- End Key Results -->
<!-- Conclusion and Takeaway Section -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<h3 class="title is-4 mt-5">Conclusion & Takeaways</h3>
<p class="content has-text-justified">
Our systematic study reveals that most current multi-encoder MLLMs exhibit <strong>substantial redundancy</strong>, especially on general-knowledge benchmarks. While multiple encoders offer some specialization—most notably for OCR and chart tasks—models can often be <strong>simplified</strong> (by masking or re-training with fewer encoders) with only a modest drop in overall accuracy. This yields significant <strong>efficiency gains</strong> in training and inference, with dual-encoder variants retaining >90% of the full model’s performance outside the most vision-centric domains.
</p>
<div class="content has-text-justified">
<strong>Takeaway:</strong>
<ul>
<li>Encoder redundancy is evident: for general benchmarks, a single encoder is sufficient; for more complex or vision-centric scenarios, two encoders suffice—adding additional encoders brings little or no improvement.</li>
<li>Language-aligned and image-trained encoders do <strong>not</strong> show any significant difference in overall results.</li>
<li>Eliminating extra encoders and re-training the model notably improves efficiency while maintaining high performance.</li>
</ul>
</div>
</section>
<!-- Video presentation: optional, add YouTube ID when available -->
<!--
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container">
<h2 class="title is-3">Video Presentation</h2>
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="publication-video">
<iframe src="https://www.youtube.com/embed/YOUR_VIDEO_ID" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
</div>
</div>
</div>
</div>
</div>
</section>
-->
<!-- Paper poster (add poster.pdf when available) -->
<!-- <section class="hero is-small is-light">
<div class="hero-body">
<div class="container">
<h2 class="title">Poster</h2>
<p class="subtitle">Conference poster will be linked here when available.</p>
<!-- <iframe src="static/pdfs/poster.pdf" width="100%" height="550" title="ICLR 2026 Poster"></iframe> -->
<!-- </div>
</div>
</section> --> -->
<!-- End paper poster -->
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div class="bibtex-header">
<h2 class="title">BibTeX</h2>
<button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX to clipboard">
<i class="fas fa-copy"></i>
<span class="copy-text">Copy</span>
</button>
</div>
<pre id="bibtex-code"><code>@inproceedings{
wang2026investigating,
title={Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders},
author={Yizhou Wang and Song Mao and Yang Chen and Yufan Shen and Pinlong Cai and Ding Wang and Guohang Yan and Zhi Yu and Yinqiao Yan and Xuming Hu and Botian Shi},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=cAopJVLKvi}
}
</code></pre>
</div>
</section>
<!--End BibTex citation -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a> which was adopted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
You are free to borrow the source code of this website, we just ask that you link back to this page in the footer. <br> This website is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Statcounter tracking code -->
<!-- You can add a tracker to track page visits by creating an account at statcounter.com -->
<!-- End of Statcounter Code -->
</body>
</html>