-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
282 lines (249 loc) · 14.8 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
<!-- Replace the content tag with appropriate information -->
<meta name="description" content="DESCRIPTION META TAG">
<meta property="og:title" content="SOCIAL MEDIA TITLE TAG"/>
<meta property="og:description" content="SOCIAL MEDIA DESCRIPTION TAG TAG"/>
<meta property="og:url" content="URL OF THE WEBSITE"/>
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="static/image/your_banner_image.png" />
<meta property="og:image:width" content="1200"/>
<meta property="og:image:height" content="630"/>
<meta name="twitter:title" content="TWITTER BANNER TITLE META TAG">
<meta name="twitter:description" content="TWITTER BANNER DESCRIPTION META TAG">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="static/images/your_twitter_banner_image.png">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="KEYWORDS SHOULD BE PLACED HERE">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Academic Project Page</title>
<link rel="icon" type="image/x-icon" href="static/images/favicon.ico">
<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">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.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/index.js"></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">Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization
</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block">
<a href="https://shuoxing98.github.io/" target="_blank">Shuo Xing</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=HqULCxoAAAAJ&hl=en" target="_blank">Yueping Wang</a><sup>2</sup>,</span>
<span class="author-block">
<a href="https://github.com/PeiranLi0930" target="_blank">Peiran Li</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/patrick-bai/" target="_blank">Ruizheng Bai</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://orcid.org/0009-0003-1374-7082" target="_blank">Yueqi Wang</a><sup>3</sup>,
</span><br>
<span class="author-block">
<a href="https://qiancx.com/" target="_blank">Chengxuan Qian</a><sup>1</sup><sup>*</sup>,
</span>
<span class="author-block">
<a href="https://www.huaxiuyao.io/" target="_blank">Huaxiu Yao</a><sup>4</sup>,
</span>
<span class="author-block">
<a href="https://vztu.github.io/" target="_blank">Zhengzhong Tu</a><sup>1</sup>,
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Texas A&M University, <sup>2</sup>University of Michigan,<br>
<sup>3</sup>University of Illinois Urbana-Champaign,
<sup>4</sup>University of North Carolina at Chapel Hill<br>
Conferance name and year</span>
<span class="eql-cntrb"><small><br><sup>*</sup>Work done during the internship at Texas A&M University</small></span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- Arxiv PDF link -->
<span class="link-block">
<a href="https://arxiv.org/pdf/<ARXIV PAPER ID>.pdf" target="_blank"
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>
<!-- Supplementary PDF link
<span class="link-block">
<a href="static/pdfs/supplementary_material.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Supplementary</span>
</a>
</span> -->
<!-- Github link -->
<span class="link-block">
<a href="https://github.com/YOUR REPO HERE" 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>
<!-- ArXiv abstract Link -->
<span class="link-block">
<a href="https://arxiv.org/abs/<ARXIV PAPER ID>" 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>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Teaser video-->
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<!-- <video poster="" id="tree" autoplay controls muted loop height="100%"> -->
<!-- Your video here -->
<!-- <source src="static/videos/banner_video.mp4" -->
<!-- type="video/mp4"> -->
<!-- </video> -->
<img src="asset\pics\radar-teaser.png" alt="MY ALT TEXT"/>
<h2 class="subtitle has-text-centered">
Benchmark performance comparison (min-max normalized).
</h2>
</div>
</div>
</section>
<!-- End teaser video -->
<!-- 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>
The emergence of large Vision Language Models (VLMs) has broadened the scope and capabilities of single-modal Large Language Models (LLMs) by integrating visual modalities, thereby unlocking transformative cross-modal applications in a variety of real-world scenarios. Despite their impressive performance, VLMs are prone to significant hallucinations, particularly in the form of cross-modal inconsistencies. Building on the success of Reinforcement Learning from Human Feedback (RLHF) in aligning LLMs, recent advancements have focused on applying direct preference optimization (DPO) on carefully curated datasets to mitigate these issues. Yet, such approaches typically introduce preference signals in a brute-force manner, neglecting the crucial role of visual information in the alignment process. In this paper, we introduce Re-Align, a novel alignment framework that leverages image retrieval to construct a dual-preference dataset, effectively incorporating both textual and visual preference signals. We further introduce rDPO, an extension of the standard direct preference optimization that incorporates an additional visual preference objective during fine-tuning. Our experimental results demonstrate that Re-Align not only mitigates hallucinations more effectively than previous methods but also yields significant performance gains in general visual question-answering (VQA) tasks. Moreover, we show that Re-Align maintains robustness and scalability across a wide range of VLM sizes and architectures. This work represents a significant step forward in aligning multimodal LLMs, paving the way for more reliable and effective cross-modal applications.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- Image carousel -->
<section class="hero is-small">
<div class="hero-body">
<div class="container is-max-desktop content">
<h2 class="title is-3">Preference Generation</h2>
<img src="asset\pics\reject-response.png" alt="MY ALT TEXT" style="display: block; margin: 0 auto; width: 800px; height: auto;"/>
<div class="content has-text-centered"><p></p><strong>Figure 1.</strong>
<em>Illustration of the preference generation process</em>,
utilizing the original vision encoder from initial VLMs
and the SentenceTransformer as the text encoder.</div>
<div class="content has-text-justified">
<p>
</p>
<ul>
<li><strong>Strategical masking: </strong>Given an input pair and its corresponding chosen response generated by a pretrained VLM, a strategic masking process removes words or segments associated with objects, attributes, or logical relationships inferred from the image, producing the masked response.</li>
<li><strong>Image retrieval: </strong>All images in the training set are embedded using the original vision encoder of the pre-trained VLMs, forming the knowledge base. The top-k most similar images to the input image are then retrieved from knowledge base using a cosine similarity search.</li>
<li><strong>Inducing hallucinations: </strong>VLMs are prompted to generate a candidate completion for the masked response conditioned on the input instruction and one of the retrieved image, which are ranked by their cosine similarity to the input image.. Both the chosen response and the reconstructed response are embedded using a <em>SentenceTransformer</em> model. If the cosine similarity between these embeddings falls below 0.95, the reconstructed response is designated as the rejected response. Otherwise, the process continues with the next retrieved image in the similarity-ranked sequence until a suitable candidate is identified or all retrieved images have been examined.
</li>
</ul>
</div>
</div>
</div>
</section>
<!-- End image carousel -->
<!-- Image carousel -->
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop content">
<h2 class="title is-3">Preference Optimization</h2>
<p>
We propose <strong>retrieval-augmented direct preference optimization (rDPO)</strong>, an extension of DPO that integrates an additional visual preference optimization objective, which is formulated as follows:
</p>
<img src="asset\pics\rDPO.png" alt="MY ALT TEXT" style="display: block; margin: 0 auto; width: 700px; height: auto;"/>
</div>
</div>
</div>
</section>
<!-- End image carousel -->
<!-- Image carousel -->
<section class="hero is-small">
<div class="hero-body">
<div class="container is-max-desktop content">
<h2 class="title is-3">Results</h2>
<p>
Re-Align achieves the best among the evaluated methods on both POPE and HallusionBench for LLaVA-v1.5-7B and LLaVA-v1.6-Mistral-7B, highlighting the effectiveness of our approach in mitigating hallucinations of VLMs. Furthermore, Re-Align can provide generally on-par or better performance than the vanilla models and baseline alignment methods on each evaluated general VQA task, ultimately achieving the best overall results.
</p>
<img src="asset\pics\table1.png" alt="MY ALT TEXT" style="display: block; margin: 0 auto; width: 800px; height: auto;"/>
<div class="content has-text-centered"><p></p><strong>Table 1. </strong>
<em>Impact of Re-Align across hallucination benchmarks for VLMs, and comparisons with baselines.</em>
</div>
<img src="asset\pics\table2.png" alt="MY ALT TEXT" style="display: block; margin: 0 auto; width: 800px; height: auto;"/>
<div class="content has-text-centered"><p></p><strong>Table 2. </strong>
<em>Impact of Re-Align across general benchmarks for VLMs, and comparisons with baselines.</em>
</div>
</div>
</div>
</div>
</section>
<!-- End image carousel -->
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>BibTex Code Here</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>