-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathCSR.html
275 lines (253 loc) · 12.7 KB
/
CSR.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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Calibrated Self-Rewarding Vision Language Models">
<meta name="keywords" content="Calibrated Self-Rewarding, Large Vision-Language Models, Modality Alignment">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta property="og:url" content="https://dongjie-cheng.github.io/CSR.html"/>
<title> Calibrated Self-Rewarding Vision Language Models</title>
<link rel="icon" href="./assets/csr.jpg">
<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 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>
</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;">
<a class="navbar-item" href="https://yiyangzhou.github.io/">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
<a class="navbar-item" href="https://x.com/AiYiyangZ/">
<span class="icon">
<i class="fab fa-twitter"></i>
</span>
</a>
<a class="navbar-item" href="https://www.linkedin.com/in/yiyang-zhou-1bb05829a">
<span class="icon">
<i class="fas fa-linkedin"></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://github.com/YiyangZhou/LURE">
<b>LURE</b> <p style="font-size:18px; display: inline; margin-left: 5px;"></p>
</a>
<a class="navbar-item" href="https://github.com/YiyangZhou/POVID">
<b>POVID</b> <p style="font-size:18px; display: inline; margin-left: 5px;"></p>
</a>
<a class="navbar-item" href="https://github.com/gzcch/Bingo">
<b>Bingo</b> <p style="font-size:18px; display: inline; margin-left: 5px;"></p>
</a>
<a class="navbar-item" href="https://github.com/X-PLUG/mPLUG-Owl">
<b>mplugOwl</b> <p style="font-size:18px; display: inline; margin-left: 5px;"></p>
</a>
<a class="navbar-item" href="https://github.com/UCSC-VLAA/vllm-safety-benchmark">
<b>Unicorn</b> <p style="font-size:18px; display: inline; margin-left: 5px;"></p>
</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">
<span>Calibrated Self-Rewarding Vision Language Models</span>
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://yiyangzhou.github.io/">Yiyang Zhou</a>*<sup style="color:#005587;">1</sup>,</span>
<span class="author-block">
<a href="https://zhiyuan.fan/">Zhiyuan Fan</a>*<sup style="color:#066221;">5</sup>,</span>
<span class="author-block">
<a href="https://dongjie-cheng.github.io/">Dongjie Cheng</a>*<sup style="color:#066221;">5</sup>,
</span>
<span class="author-block">
<a href="https://dongjie-cheng.github.io/CSR.html">Sihan Yang</a><sup style="color:#066221">5</sup>,
</span>
<span class="author-block">
<a href="https://billchan226.github.io/">Zhaorun Chen</a><sup style="color:#FDB515">2</sup>,
</span>
<span class="author-block">
<a href="https://gzcch.github.io/">Chenhang Cui</a><sup style="color:#066221">5</sup>,
</span>
<span class="author-block">
<a href="https://si0wang.github.io/">Xiyao Wang</a><sup style="color:#8C1515;">3</sup>,
</span>
<span class="author-block">
<a href="https://yunliweb.its.unc.edu/people.html#YunLi">Yun Li</a><sup style="color:#005587">1</sup>
</span>
<span class="author-block">
<a href="https://linjunz.github.io/">Linjun Zhang</a><sup style="color:#d7237d">4</sup>
</span>
<span class="author-block">
<a href="https://sites.google.com/view/danicaxiao/home">Huaxiu Yao</a><sup style="color:#005587">1</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup style="color:#005587;">1</sup>UNC-Chapel Hill,</span><br>
<span class="author-block"><sup style="color:#FDB515">2</sup>University of Chicago,</span>
<span class="author-block"><sup style="color:#8C1515">3</sup>University of Maryland,</span>
<span class="author-block"><sup style="color:#d7237d">4</sup>Rutgers University,</span>
<span class="author-block"><sup style="color:#066221">5</sup>Independent Researcher</span><br>
<span class="paper-block"><b >* Equal Contribution</b> </span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/pdf/2405.14622"
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/2405.14622"
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/YiyangZhou/CSR"
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/charlesdj"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<p style="font-size:18px">🤗</p>
</span>
<span>Hugging Face</span>
</a>
</span>
<!-- Twitter Link. -->
<span class="link-block">
<a href="https://x.com/HuaxiuYaoML/status/1794203052116680895"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<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="assets/framework.png" alt="geometric reasoning" width="100%"/>
<p>
The CSR framework operates an iterative process of preference data generation and learning. During preference data generation, CSR utilizes a sentence-level beam search approach to construct responses sentence by sentence, assigning a reward to each sentence. This reward, initially generated by the model itself, is then calibrated using image-relevance information. Preferences are determined based on the cumulative reward for each response. In each iteration, CSR generates new preference data and performs preference learning based on this data, continuously enhancing the model's performance.
</p>
</div>
</div>
</section>
<!-- Introduction -->
<section class="section">
<div class="container" style="margin-bottom: 2vh;">
<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">
Existing methods use additional models or human annotations to curate preference data and enhance modality alignment through preference optimization. These methods are resource-intensive and may not effectively reflect the target LVLM's preferences, making the curated preference data easily distinguishable. To address these challenges, we proposes the <b>C</b>alibrated <b>S</b>elf-<b>R</b>ewarding (<b>CSR</b>), which enables the model to self-improve by iteratively generating candidate responses, evaluating the reward for each response, and curating preference data for fine-tuning. In reward modeling, a step-wise strategy is adopted, and visual constraints are incorporated into the self-rewarding process to emphasize visual input.
</div>
<p align="center">
<img src="assets/csr_llava.png" width="90%"> <br>
Left: Different parameter sizes of LLaVA 1.5 can enhance their learning through CSR iterations. Right: The change in image relevance scores before and after employing CSR.
</p>
<div class="content has-text-justified">
Through the online CSR process, the model continuously enhances its performance across various benchmarks and improves the overall relevance scores of its responses to visual inputs. Additionally, it reduces the gap between rejected responses and chosen responses, thereby improving the model's performance lower bound.
</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>@article{zhou2024calibrated,
title={Calibrated Self-Rewarding Vision Language Models},
author={Zhou, Yiyang and Fan, Zhiyuan and Cheng, Dongjie and Yang, Sihan and Chen, Zhaorun and Cui, Chenhang and Wang, Xiyao and Li, Yun and Zhang, Linjun and Yao, Huaxiu},
journal={arXiv preprint arXiv:2405.14622},
year={2024}
}</code></pre>
</div>
</section>
<section>
<div class="section" id="org-banners" style="display:flex">
<a href="https://www.unc.edu/" target="_blank" rel="external">
<img class="center-block org-banner" src="assets/UNC-Logo.png" style="max-width: 200px; height: auto;">
</a>
<a href="https://www.uchicago.edu/" target="blank" class="ext-link">
<img class="center-block org-banner" src="assets/uchi_logo.png" style="max-width: 200px; height: auto;">
</a>
<a href="https://umd.edu/" target="_blank" rel="external">
<img class="center-block org-banner" src="assets/University-of-Maryland-Logo.png" style="max-width: 200px; height: auto;">
</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>, 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>