-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
322 lines (287 loc) · 16.6 KB
/
index.html
File metadata and controls
322 lines (287 loc) · 16.6 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<br>
<div class="logo" style="text-align: center;">
<a href="index.html">
<img src="./assets/images/logo.webp" width="10%">
</a>
</div>
<title>Towards Tracing Trustworthiness Dynamics: <br>Revisiting Pre-training Period of Large Language Models</title>
<script>
var task_map = {
"simple-object-manipulation": "simple_object_manipulation",
"visual-goal-reaching": "visual_goal_reaching",
"novel-concept-grounding": "novel_concept_grounding",
"one-shot-video-imitation": "one_shot_video_imitation",
"visual-constraint-satisfaction": "visual_constraint_satisfaction",
"visual-reasoning": "visual_reasoning"
};
function updateDemoVideo(category) {
// var demo = document.getElementById("single-menu-demos").value;
var task = document.getElementById(category + "-menu-tasks").value;
var inst = document.getElementById(category + "-menu-instances").value;
console.log(task_map[category], task, inst)
var video = document.getElementById(category + "-single-task-video");
video.src = "assets/videos/demos/" +
task_map[category] +
"/" +
task +
"/" +
inst +
".mp4";
video.playbackRate = 2.0;
video.play();
}
</script>
<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="./static/css/academicons.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/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" style="font-size: 2.5rem;">Towards Tracing Trustworthiness Dynamics: <br>Revisiting Pre-training Period of Large Language Models</h1>
<h3 class="title is-4 conference-authors"><a target="_blank" href="https://cvpr.thecvf.com/">ACL 2024</a>
</h3>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a target="_blank" href="https://github.com/ChnQ">Chen Qian</a><sup>1 2*</sup>,
<a target="_blank" href="https://github.com/tmylla">Jie Zhang</a><sup>1 3*</sup>,
<a target="_blank"
href="https://scholar.google.com/citations?user=JbYPAqoAAAAJ&hl=en">Wei Yao</a><sup>1 2*</sup>,
<a target="_blank" href="https://scholar.google.com/citations?user=i0paeq4AAAAJ&hl=en">Dongrui Liu</a><sup>1 4</sup>,
<br>
<a target="_blank" href="https://scholar.google.com/citations?user=ngPR1dIAAAAJ&hl=en">Zhenfei Yin</a><sup>1 5</sup>,
<a target="_blank"
href="https://scholar.google.com/citations?user=gFtI-8QAAAAJ&hl=en">Yu Qiao</a><sup>1</sup>,
<a target="_blank" href="https://scholar.google.com/citations?hl=en&user=vVhmzbAAAAAJ">Yong Liu</a><sup>2✉</sup>,
<a target="_blank" href="https://amandajshao.github.io/">Jing Shao</a><sup>1✉</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Shanghai Artificial Intelligence Laboratory; </span>
<span class="author-block"><sup>2</sup>Renmin University of China; </span>
<span class="author-block"><sup>3</sup>Chinese Academy of Sciences; </span>
<span class="author-block"><sup>4</sup>Shanghai Jiao Tong University; </span>
<span class="author-block"><sup>5</sup>The University of Sydney; </span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>* </sup>Equal Contribution  </span>
<span class="author-block"><sup>✉ </sup>Corresponding author  </span>
<!-- <span class="author-block"><sup>† </sup>Project Leader  </span> -->
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- TODO PDF Link. -->
<span class="link-block">
<a target="_blank" href="https://arxiv.org/abs/2402.19465"
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 target="_blank" href="assets/TracingLLMs.pdf"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>PDF</span>
</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a target="_blank" href="https://github.com/ChnQ/TracingLLM"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
<a target="_blank" href="https://github.com/ChnQ/TracingLLM/tree/master/src/datasets"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-database"></i>
</span>
<span>Dataset</span>
</a>
</span>
<span class="link-block">
<a target="_blank" href="https://mp.weixin.qq.com/s/JD0PSOahuYloTgOHqGUtcA"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-book"></i>
</span>
<span>WeChat Blog</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<div class="columns is-centered has-text-centered">
<div class="column">
<img src="assets/images/overview.png" width="75%">
</div>
</div>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column">
<h2 class="title is-3">Introduction</h2>
<div class="content has-text-justified">
<p style="font-size: 125%">
We are excited to present "Towards Tracing Trustworthiness Dynamics: Revisiting Pre-training Period of Large Language Models," a pioneering study on exploring trustworthiness in LLMs during pre-training.
We explores five key dimensions of trustworthiness: reliability, privacy, toxicity, fairness, and robustness.
By employing linear probing and extracting steering vectors from LLMs' pre-training checkpoints, the study aims to uncover the potential of pre-training in enhancing LLMs' trustworthiness. Furthermore, we investigates the dynamics of trustworthiness during pre-training through mutual information estimation, observing a two-phase phenomenon: fitting and compression.
Our findings unveil new insights and encourage further developments in improving the trustworthiness of LLMs from an early stage.
</p>
</div>
</div>
</div>
</div>
</section>
<!--Model-->
<!--Model-->
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered ">
<div class="row is-full-width">
<h2 class="title is-3"><span class="dvima">Probing LLM Pre-training Dynamics in Trustworthiness</span></h2>
<img src="assets/images/probing.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<br>
<span style="font-size: 110%">
<span style="font-weight: bold">The linear probe accuracy on five trustworthiness dimensions for the first 80 pre-training checkpoints.</span>
<ul style="list-style-type: disc; padding-left: 20px;">
<li>Models during the <span style="color: orange; font-weight: bold;">early stages of pre-training</span> can already encode trustworthiness well.</li>
<li><span style="color: orange; font-weight: bold;">Middle and high layers</span> representations exhibit relatively high linearly separable patterns about trustworthiness than low layers.</li>
</ul>
</span>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered ">
<div class="row is-full-width">
<h2 class="title is-3"><span class="dvima">Controlling Trustworthiness with the Help of Pre-training Checkpoints </span></h2>
<div class="columns is-vcentered">
<div class="column has-text-centered">
<img src="assets/images/steering_vector.png" class="interpolation-image" width="85%"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<p style="font-size: 90%; font-weight: bold;">Constructing steering vectors from the pre-training checkpoints <br>and intervening in the SFT model</p>
</div>
<div class="column has-text-centered">
<img src="assets/images/enhancing_trustworthiness.png" class="interpolation-image" width="70%"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<p style="font-size: 90%; font-weight: bold;">Performance of various models across four general capabilities <br>and five trustworthiness capabilities</p>
</div>
</div>
<br>
<span style="font-size: 110%">
<span style="font-weight: bold">When using the steering vector from the pre-training checkpoints to intervene in the SFT model: </span>
<ul style="list-style-type: disc; padding-left: 20px;">
<li>There is a significant <span style="color: orange; font-weight: bold;">improvement</span> in three dimensions of trustworthiness.</li>
<li>The intervention <span style="color: orange; font-weight: bold;">has a minor impact on the general capabilities</span> of the model.</li>
<li>Enhance the trustworthiness performance of the SFT model <span style="color: orange; font-weight: bold;">more effectively compared to the steering vectors from the SFT model itself</span>.</li>
</ul>
</span>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered ">
<div class="row is-full-width">
<h2 class="title is-3"><span class="dvima">Probing LLMs using Mutual Information </span></h2>
<img src="assets/images/mi.png" class="interpolation-image" width="60%"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<br>
<span style="font-size: 110%">
<ul style="list-style-type: disc; padding-left: 20px;">
<li>We take an alternative view by probing LLMs with mutual information during pre-training.</li>
<li>During the pre-training period of LLMs, there exist two distinct phases regarding trustworthiness: <span style="color: orange; font-weight: bold;">fitting and compression</span>, which is in line with previous research on traditional DNNs.</li>
</ul>
</span>
</div>
</div>
</div>
</div>
</section>
<!--Conclusion-->
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered ">
<div class="row is-full-width">
<h2 class="title is-3"><span
class="dvima">Conclusion</span></h2>
<p style="font-size: 125%">
In this work, we take an initial and illuminating step towards elucidating the conceptual understanding of trustworthiness during pre-training.
Firstly, by linear probing LLMs across reliability, privacy, toxicity, fairness, and robustness, we investigate the ability of LLMs representations to discern opposing concepts within each trustworthiness dimension during the whole pre-training period.
Furthermore, motivated by the probing results, we conduct extensive experiments to reveal the potential of utilizing representations from LLMs during its previous pre-training period to enhance LLMs' own trustworthiness.
Finally, we use mutual information to probe LLMs during pre-training and reveal some similarities in the learning mechanism between LLMs and traditional DNNs.
<br><br>
Taken collectively, the empirical study presented in this work can not only justify the potential to improve the trustworthiness of LLMs using their own pre-training checkpoints but may also lead to a better understanding of the dynamics of LLM representations, especially the trustworthiness-related concepts.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- <section class="section" id="BibTeX">
<div class="container is-max-widescreen content">
<h2 class="title">BibTeX</h2>
<pre><code>@inproceedings{jiang2023vima,
title = {VIMA: General Robot Manipulation with Multimodal Prompts},
author = {Yunfan Jiang and Agrim Gupta and Zichen Zhang and Guanzhi Wang and Yongqiang Dou and Yanjun Chen and Li Fei-Fei and Anima Anandkumar and Yuke Zhu and Linxi Fan},
booktitle = {Fortieth International Conference on Machine Learning},
year = {2023}
}</code></pre>
</div>
</section> -->
<section class="section" id="BibTeX">
<div class="container is-max-widescreen content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{qian2024towards,
title={Towards Tracing Trustworthiness Dynamics: Revisiting Pre-training Period of Large Language Models},
author={Qian, Chen and Zhang, Jie and Yao, Wei and Liu, Dongrui and Yin, Zhenfei and Qiao, Yu and Liu, Yong and Shao, Jing},
journal={arXiv preprint arXiv:2402.19465},
year={2024}
}</code></pre>
</div>
</section>
</body>
</html>