forked from PKU-Aligner/PKU-Aligner.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
1531 lines (1375 loc) · 86.1 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
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<script src="bootstrap.js"></script>
<script type="text/javascript" charset="utf-8" src="https://ajax.googleapis.com/ajax/libs/jquery/1.3.2/jquery.min.js"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.1/css/all.min.css">
<!-- <link href="https://fonts.googleapis.com/icon?family=Material+Icons+Outlined" rel="stylesheet"> -->
<script>
document.addEventListener("DOMContentLoaded", function() {
// 查找所有 cite 元素
var cites = document.querySelectorAll("cite");
var footnoteMap = {}; // 用于跟踪脚注的映射
cites.forEach(function(cite, index) {
var citeContent = cite.innerHTML.trim();
var footnoteId;
if (footnoteMap[citeContent]) {
footnoteId = footnoteMap[citeContent];
}else{
// 创建新的脚注ID
footnoteId = (Object.keys(footnoteMap).length + 1);
footnoteMap[citeContent] = footnoteId;
// 创建新的脚注内容
var footnote = document.createElement("li");
footnote.innerHTML = "<b>[" + (Object.keys(footnoteMap).length) + "]</b> " + citeContent;
footnote.id = footnoteId;
document.getElementById("footnotes-list").appendChild(footnote);
}
// 在原位置添加引用编号
var sup = document.createElement("sup");
var a = document.createElement("a");
a.href = "#" + footnoteId;
a.innerHTML = "[" + footnoteId + "]";
sup.appendChild(a);
cite.parentNode.insertBefore(sup, cite.nextSibling);
// 隐藏原始的 cite 标签
cite.style.display = "none";
});
});
</script>
<!-- <script>
function onScroll() {
var sections = document.querySelectorAll("section");
var menuLinks = document.querySelectorAll("#table-of-contents");
var currentSectionId = "";
sections.forEach((section) => {
var rect = section.getBoundingClientRect();
// 检查section的顶部或底部是否在视窗内
if (rect.top <= 50 && rect.bottom >= 50) {
currentSectionId = section.id;
}
});
// 更新链接的活跃状态
menuLinks.forEach((link) => {
if (link.getAttribute("href") === "#" + currentSectionId) {
link.classList.add("active-link");
} else {
link.classList.remove("active-link");
}
});
}
// 监听滚动事件
window.addEventListener("scroll", onScroll);
</script> -->
<script>
function detectDeviceType() {
const userAgent = navigator.userAgent.toLowerCase();
if (/mobile/i.test(userAgent)) {
return 'Mobile';
} else {
return 'Desktop';
}
}
console.log(detectDeviceType());
</script>
<!-- TBD -->
<style type="text/css">
body {
font-family: "Titillium Web", "HelveticaNeue-Light", "Helvetica Neue Light", "Helvetica Neue", Helvetica, Arial, "Lucida Grande", sans-serif;
font-weight: 300;
font-size: 17px;
margin-left: auto;
margin-right: auto;
}
@media screen and (min-width: 980px){
body {
width: 980px;
}
}
h1 {
font-weight:300;
line-height: 1.15em;
}
h2 {
font-size: 1.75em;
}
a:link,a:visited {
color: #5364cc;
text-decoration: none;
}
a:hover {
color: #208799;
}
h1 {
text-align: center;
}
h2,h3 {
text-align: left;
}
h1 {
font-size: 40px;
font-weight: 500;
}
h2 {
font-weight: 400;
margin: 16px 0px 4px 0px;
}
h3 {
font-weight: 600;
margin: 16px 0px 4px 0px;
}
.paper-title {
padding: 1px 0px 1px 0px;
}
section {
margin: 32px 0px 32px 0px;
text-align: justify;
clear: both;
}
.col-5 {
width: 20%;
float: left;
}
.col-4 {
width: 25%;
float: left;
}
.col-3 {
width: 33%;
float: left;
}
.col-2 {
width: 50%;
float: left;
}
.col-1 {
width: 100%;
float: left;
}
.author-row, .affil-row {
font-size: 26px;
}
.author-row-new {
text-align: center;
}
.author-row-new a {
display: inline-block;
font-size: 20px;
padding: 4px;
}
.author-row-new sup {
color: #313436;
font-size: 12px;
}
.affiliations-new {
font-size: 18px;
text-align: center;
width: 80%;
margin: 0 auto;
margin-bottom: 20px;
}
.row {
margin: 16px 0px 16px 0px;
}
.authors {
font-size: 26px;
}
.affiliatons {
font-size: 18px;
}
.affil-row {
margin-top: 18px;
}
.teaser {
max-width: 100%;
}
.text-center {
text-align: center;
}
.screenshot {
width: 256px;
border: 1px solid #ddd;
}
.screenshot-el {
margin-bottom: 16px;
}
hr {
height: 1px;
border: 0;
border-top: 1px solid #ddd;
margin: 0;
}
.material-icons {
vertical-align: -6px;
}
.material-symbols-outlined{
vertical-align: -6px;
}
p {
line-height: 1.25em;
}
.caption {
font-size: 16px;
color: #666;
margin-top: 4px;
margin-bottom: 10px;
}
video {
display: block;
margin: auto;
}
details { background-color: #f0f0f0; } /* 浅灰色背景 */
summary { background-color: #FFFFFF; }
figure {
display: block;
margin: auto;
margin-top: 10px;
margin-bottom: 10px;
}
#bibtex pre {
font-size: 14px;
background-color: #eee;
padding: 16px;
}
.blue {
color: #2c82c9;
font-weight: bold;
}
.orange {
color: #d35400;
font-weight: bold;
}
.flex-row {
display: flex;
flex-flow: row wrap;
padding: 0;
margin: 0;
list-style: none;
}
.paper-btn-coming-soon {
position: relative;
top: 0;
left: 0;
}
.coming-soon {
position: absolute;
top: -15px;
right: -15px;
}
.paper-btn {
position: relative;
text-align: center;
display: inline-block;
margin: 8px;
padding: 8px 8px;
border-width: 0;
outline: none;
border-radius: 2px;
background-color: #5364cc;
color: white !important;
font-size: 20px;
width: 100px;
font-weight: 600;
}
.paper-btn-parent {
display: flex;
justify-content: center;
margin: 16px 0px;
}
.paper-btn:hover {
opacity: 0.85;
}
.container {
margin-left: auto;
margin-right: auto;
padding-left: 16px;
padding-right: 16px;
}
.venue {
font-size: 23px;
}
.topnav {
background-color: #EEEEEE;
overflow: hidden;
}
.topnav div {
max-width: 1070px;
margin: 0 auto;
}
.topnav a {
display: inline-block;
color: black;
text-align: center;
vertical-align: middle;
padding: 16px 16px;
text-decoration: none;
font-size: 18px;
}
.topnav img {
padding: 2px 0px;
width: 100%;
margin: 0.2em 0px 0.3em 0px;
vertical-align: middle;
}
pre {
font-size: 0.9em;
padding-left: 7px;
padding-right: 7px;
padding-top: 3px;
padding-bottom: 3px;
border-radius: 3px;
background-color: rgb(235, 235, 235);
overflow-x: auto;
}
.download-thumb {
display: flex;
}
@media only screen and (max-width: 620px) {
.download-thumb {
display: none;
}
}
.paper-stuff {
width: 50%;
font-size: 20px;
}
@media only screen and (max-width: 620px) {
.paper-stuff {
width: 100%;
}
}
* {
box-sizing: border-box;
}
.column {
text-align: center;
float: left;
width: 16.666%;
padding: 5px;
}
.column3 {
text-align: center;
float: left;
width: 33.333%;
padding: 5px;
}
.column4 {
text-align: center;
float: left;
width: 50%;
padding: 5px;
}
.column5 {
text-align: center;
float: left;
width: 20%;
padding: 5px;
}
.border-right {
border-right: 1px solid black;
}
.border-bottom{
border-bottom: 1px solid black;
}
/* Clearfix (clear floats) */
.row::after {
content: "";
clear: both;
display: table;
}
.img-fluid {
max-width: 100%;
height: auto;
}
.figure-img {
margin-bottom: 0.5rem;
line-height: 1;
}
.pdf-object {
border: none; /* 移除边框 */
}
.center-content {
text-align: center; /* 确保内容居中 */
}
.center-content img {
margin: 0 auto; /* 水平居中 */
display: block; /* 使 img 行为像块级元素 */
}
.rounded-circle {
border-radius: 50% !important;
}
.red-bold-italic {
color: red; /* 设置文字颜色为红色 */
font-weight: bold; /* 设置文字为加粗 */
font-style: italic;/* 设置文字为斜体 */
}
#teaser-image {
text-align: center; /* 确保内容居中 */
}
#teaser-image figure {
display: inline-block; /* 使 figure 行为像内联元素，以便居中 */
margin: 0 auto; /* 水平居中 */
}
.image-container {
display: flex;
flex-direction: row; /* 设置为行布局，使图像水平排列 */
justify-content: space-between; /* 在容器中水平分散排列图像 */
align-items: center; /* 在容器中垂直居中对齐图像 */
}
.image-container br {
margin-top: 20px; /* 调整间距的大小，根据需要进行调整 */
}
#table-of-contents {
position: fixed; /* 固定位置 */
left: 20px; /* 距离窗口右边缘20px */
top: 50px; /* 距离窗口顶部100px */
width: 200px; /* 目录的宽度 */
background-color: white; /* 背景色 */
padding: 10px; /* 内边距 */
border: 1px solid #ddd; /* 边框样式 */
box-shadow: 0px 0px 10px #ddd; /* 阴影效果 */
z-index: 1000; /* 确保目录在最顶层 */
}
#table-of-contents h2 {
margin-top: 0; /* 移除标题的顶部外边距 */
}
#table-of-contents ul {
list-style-type: none; /* 去除列表符号 */
padding: 0; /* 去除内边距 */
}
#table-of-contents li {
margin-bottom: 10px; /* 列表项之间的间距 */
}
#table-of-contents a {
text-decoration: none; /* 移除链接下划线 */
color: #000; /* 设置链接颜色 */
}
#table-of-contents a:hover {
text-decoration: underline; /* 鼠标悬停时添加下划线 */
}
.active-link {
color: blue; /* 设置活跃链接的颜色为蓝色 */
}
/* Responsive layout - makes the three columns stack on top of each other instead of next to each other */
@media screen and (max-width: 500px) {
.column {
width: 100%;
}
}
@media screen and (max-width: 500px) {
.column3 {
width: 100%;
}
}
</style>
<link rel="stylesheet" href="bootstrap-grid.css">
<script type="text/javascript" src="../js/hidebib.js"></script>
<link href='https://fonts.googleapis.com/css?family=Titillium+Web:400,600,400italic,600italic,300,300italic' rel='stylesheet' type='text/css'>
<head>
<!-- <title> Achieving Efficient Alignment through Learned Correction</title> -->
<title>Aligner Paper Website</title>
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta property="og:description" content="Aligner: Achieving Efficient Alignment through Answer Correction"/>
<link href="https://fonts.googleapis.com/css2?family=Material+Icons" rel="stylesheet">
<meta name="twitter:title" content="Aligner: Achieving Efficient Alignment through Answer Correction">
<meta name="twitter:description" content="">
<meta name="twitter:image" content="">
</head>
<body>
<div class="container">
<div class="paper-title">
<h1>
<!-- <i>Aligner</i> :
Achieving Efficient Alignment through <br>
Weak-to-Strong Correction -->
<!-- Achieving Efficient Alignment through Learned Correction -->
<style>
.paper-title h1 {
font-family: 'Times New Roman', Times, serif;
font-weight: bold; /* Make the text bold */
font-size: 100px; /* Set the font size */
margin-bottom: 10px; /* Reduce the bottom margin */
}
</style>
<div class="paper-title">
<h1>ALIGNER</h1>
</div>
<!-- <figure class="center-content">
<img src="materials/logo2.png" style="width: 100%;">
</figure> -->
</h1>
</div>
<div id="authors">
<!-- <center>
<div class="author-row-new"> -->
<!-- <a>Anonymous Authors*</a> -->
<!-- <a href="https://anuragajay.github.io/">Anurag Ajay*<sup>1,2</sup></a>,
<a href="https://yilundu.github.io/">Yilun Du*<sup>2</sup></a>,
<a href="https://scholar.google.com/citations?user=ynyPc1kAAAAJ&hl=en">Abhi Gupta*<sup>2</sup></a>,
<a href="https://scholar.google.com/citations?user=rRJ9wTJMUB8C&hl=en">Joshua B. Tenenbaum<sup>2</sup></a>,
<a href="http://people.csail.mit.edu/tommi/">Tommi Jaakkola<sup>2</sup></a>,
<a href="https://people.csail.mit.edu/pulkitag/">Pulkit Agrawal<sup>1,2</sup></a> -->
<!-- </div>
</center>
<center> -->
<!-- <div class="affiliations">
<span><sup>1</sup> Improbable AI Lab</span>
<span><sup>2</sup> MIT</span><br/>
</div> -->
<!-- <br>*under review -->
<!-- <div class="affil-row">
<div class="venue text-center"><b>ICML 2024 <FONT COLOR="RED">(Under Review)</FONT></b></div>
</div> -->
<!-- </center> -->
<div style="clear: both">
<div class="paper-btn-parent">
<a class="paper-btn" href="materials/Aligner_anonymous.pdf" target="_blank">
<span class="material-icons"> description </span>
Paper
</a>
<div class="paper-btn-coming-soon">
<a class="paper-btn" href="https://github.com/PKU-Alignment/aligner" target="_blank">
<span class="material-icons"> code </span>
Code
</a>
</div>
<div class="paper-btn-coming-soon">
<a class="paper-btn" href="https://huggingface.co/datasets/aligner/aligner-20K" target="_blank">
<span class="material-icons"> folder </span>
Data
</a>
</div>
<div class="paper-btn-coming-soon">
<a class="paper-btn" href="https://huggingface.co/aligner/aligner-7b-v1.0" target="_blank">
<span class="material-icons"> download </span>
Model
</a>
</div>
<div class="paper-btn-coming-soon">
<a class="paper-btn" href="#qa">
<span class="material-icons"> quiz </span>
FAQ
</a>
</div>
</div></div>
</div>
<!-- <section id="teaser-image">
<center>
<figure>
<video class="centered" width="80%" autoplay loop muted playsinline class="video-background " >
<source src="assets/LION_video_v10.mp4#t=0.001" type="video/mp4">
Your browser does not support the video tag.
</video>
</figure>
</center>
</section>
-->
<div id="table-of-contents">
<h2>Contents</h2>
<ul>
<li><a href="#abstract">Abstract</a></li>
<li><a href="#teaser-image">Paradigm</a></li>
<li><a href="#method">Structure & Methodology</a></li>
<li><a href="#interp">Interpretability</a></li>
<li><a href="#application">Applications</a></li>
<li><a href="#results">Results Overview</a></li>
<li><a href="#detail">Detailed Results</a></li>
<li><a href="#qa">FAQ</a></li>
<!-- 其他部分的链接 -->
</ul>
</div>
<br>
<section id="abstract"/>
<hr>
<h2>Abstract</h2>
<div class="flex-row">
<p>
With the rapid development of large language models (LLMs) and ever-evolving practical requirements, finding an efficient and effective alignment method has never been more critical. However, the tension between the complexity of current alignment methods and the need for rapid iteration in deployment scenarios necessitates the development of a model-agnostic alignment approach that can operate under these constraints. In this paper, we introduce <i>Aligner</i>, a novel and simple alignment paradigm that learns the correctional residuals between preferred and dispreferred answers using a small model. Designed as a model-agnostic, plug-and-play module, <i>Aligner</i> can be directly applied to various open-source and API-based models with only one-off training, making it suitable for rapid iteration. Notably, <i>Aligner</i> can be applied to any powerful, large-scale upstream models. Moreover, it can even iteratively bootstrap the upstream models using corrected responses as synthetic human preference data, breaking through the model's performance ceiling. Our experiments demonstrate performance improvements by deploying the same <i>Aligner</i> model across 11 different LLMs, evaluated on the 3H dimensions (helpfulness, harmlessness, and honesty). Specifically, <i>Aligner</i>-7B has achieved an average improvement of \(68.9\%\) in helpfulness and \(23.8\%\) in harmlessness across the tested LLMs while also effectively reducing hallucination. In the Alpaca-Eval leaderboard, stacking <i>Aligner</i>-2B on GPT-4 Turbo improved its LC Win Rate from \(55.0\%\) to \(58.3\%\), surpassing GPT-4 Omni's \(57.5\%\) Win Rate (community report).
</p>
</div>
</section>
<hr>
<section id="teaser-image">
<h2>Paradigm</h2>
<figure class="mx-auto" style="display: flex; justify-content: center; align-items: center;">
<img src="materials/Fig1_nocap.jpg" style="width: 550px; margin-right: 50px;">
<center><figcaption style="text-align: left;">
<strong>Architecture of the <i>Aligner</i> module and illustration of its behavior in semantic space.</strong>
<br>
The <i>Aligner</i>, a plug-and-play model <span class="red-bold-italic" >(without RLHF)</span>, stacks upon an upstream LLM (aligned or unaligned). It redistributes initial answers from the upstream model into more helpful and harmless answers, thus aligning the composed LLM responses with human intentions. It is challenging to learn direct mappings from queries to aligned answers. Nonetheless, correcting answers based on the upstream model's output is a more tractable learning task.
</figcaption></center>
</figure>
<br>
<hr>
<br>
<figure class="mx-auto" style="display: flex; justify-content: center; align-items: center;">
<img src="materials/fast-aligned-clipped.gif" style="width: 550px; margin-right: 50px;">
<center><figcaption style="text-align: left;">
<strong>Analogy of the <i>Aligner</i> as a residual learning enhancer for LLMs in both architecture and capability aspects.</strong>
<br>
This schematic showcases the <i>Aligner</i> acting similarly to a residual block in neural networks.
It takes an initial output \( y_o \) from the upstream LLM, then the <i>Aligner</i> applies its autoregressive capabilities to generate a corrected version \( y_c \).
Just as a residual block uses a shortcut to add modifications without changing the base structure, the <i>Aligner</i> employs a ''copy and correct'' method, overlaying improvements onto the original answer without altering its fundamental structure.
This parallel highlights the <i>Aligner</i>'s dual role in preserving the initial response while enhancing it to better align with desired outcomes.
</figcaption></center>
</figure>
<!-- <center> -->
<!-- <center><p><b>A unified framework for composing pre-trained models.</b></p></center>-->
<!-- <figure class="mx-auto">
<center><img class="card-img-top" src="./materials/Presentation1.jpg" width="950"/></center>
<center><figcaption style="text-align: left;">
<strong>Left: Architecture of the <i>Aligner</i> module and illustration of its behavior in semantic space.</strong>
<br>
The <i>Aligner</i>, a plug-and-play model, stacks upon an upstream LLM (aligned or unaligned). It redistributes initial answers from the upstream model into more helpful and harmless answers, thus aligning the composed LLM responses with human intentions. It is challenging to learn direct mappings from queries to aligned answers. Nonetheless, correcting answers based on the upstream model’s output is a more tractable learning task.
<br>
<strong>Right: Analogy of the <i>Aligner</i> as a residual learning enhancer for LLMs in both architecture and capability aspects.</strong>
<br>
This schematic showcases the <i>Aligner</i> acting similarly to a residual block in neural networks.
It takes an initial output \( y_o \) from the upstream LLM, then the <i>Aligner</i> applies its autoregressive capabilities to generate a corrected version \( y_c \).
Just as a residual block uses a shortcut to add modifications without changing the base structure, the <i>Aligner</i> employs a ''copy and correct'' method, overlaying improvements onto the original answer without altering its fundamental structure.
This parallel highlights the <i>Aligner</i>'s dual role in preserving the initial response while enhancing it to better align with desired outcomes.
</figcaption></center>
</figure> -->
<!-- <video width="650" loop autoplay muted>-->
<!-- <source src="materials/teaser-2.mp4" type="video/mp4">-->
<!-- </video>-->
<!-- <br><br>-->
<!-- -->
<!-- <video width="960" loop autoplay muted>-->
<!-- <source src="materials/all_results.mp4" type="video/mp4">-->
<!-- </video>-->
<!-- <br><br> -->
<!-- <video width="800" loop autoplay muted style="border:1px solid black">
<source src="materials/new3.mp4" type="video/mp4">
</video>
-->
<br>
<hr>
<br>
<figure class="mx-auto">
<center><img class="card-img-top" src="./materials/Tab3_nocap.jpg" width="950"/></center>
<center><figcaption style="text-align: left;">
<b>Performance of <i>Aligner</i> Models</b><br>
It is shown that <i>Aligner</i> achieves significant performances in all the settings. All assessments in this table were conducted based on integrating various models with Aligners to compare with the original models to quantify the percentage increase in the <i>3H</i> standard. When integrated and assessed in conjunction with various upstream models, the <i>Aligner</i> requires only a single training session (<i>i.e.</i>, the <i>Aligner</i> can operate in a zero-shot manner and enhance the performance of all upstream models.)
</figcaption></center>
</figure>
<!-- </center> -->
</section>
<section id="method"/>
<hr>
<h2>Overview of <i>Aligner</i> Methodologies</h2>
<br><br>
<!-- <figure class="mx-auto">
<center><img class="card-img-top" src="./materials/paradiam-new2.jpg" style="width:950px"></center>
<br>
<center><figcaption style="text-align: left;">
<b>An illustration of our methodology.</b><br>
The Superalignment problem focuses on scaling human oversight for supervising increasingly intelligent and complex AI systems. The <i>Weak-to-Strong Generalization</i>
<cite>
Weak-to-strong generalization: Eliciting strong capabilities with weak supervision.
</cite>
analogy emphasizes using weaker models to supervise stronger ones. Our approach composes weak and strong models to offer iteratively scalable supervision.
</figcaption></center>
</figure> -->
<!--
<hr>
<br> -->
<figure class="mx-auto">
<!-- <left><p>The proposed framework that composes a "generator" and an ensemble of "scorers" through iterative consensus enables zero-shot generalization across a variety of multimodal tasks.</p></left>-->
<center><img class="card-img-top" src="materials/table_1.jpg" style="width:950px"></center>
<br>
<center><figcaption style="text-align: left;">
<!-- <b>Overview of Alignment Methodologies</b><br> -->
The <i>Aligner</i> module, noted for its flexibility, is not constrained by specific model parameters or configurations. In contrast, traditional methods such as RLHF are limited by their need for direct access to a model's parameters. With the growth of model sizes, such as those with over 70B parameters
<cite>
Llama 2: Open foundation and fine-tuned chat models.
</cite>
, RLHF's computational demands have increased. Filter-based methods often overcorrect when replacing unsafe responses with refusals, sometimes eliminating even the safe parts of the response. An alternative approach combines both user prompts and model responses to moderation filtering
<cite>
Beavertails: Towards improved safety alignment of LLM via a human-preference dataset.
</cite>
; however, it also depends on the model’s ability to generate safe responses.
</figcaption></center>
</figure>
<br>
<!-- <div class="mx-auto"> -->
<!-- <left><p>The proposed framework that composes a "generator" and an ensemble of "scorers" through iterative consensus enables zero-shot generalization across a variety of multimodal tasks.</p></left>-->
<!-- </div> -->
<!-- <div class="row">
<div class="column4">
<center><img class="card-img-top" src="materials/framework.png" style="width:400px"></center>
</div>
<div class="column4">
<video width="400" loop autoplay muted>
<source src="materials/teaser3.mp4" type="video/mp4">
</video>
</div>
</div>
-->
<!-- <div class="flex-row">-->
<!-- <div class="mx-auto">-->
<!-- <left><p><b>Overview of the proposed unified framework.</b> Dashed lines are omitted for certain tasks. Orange lines represent the components used to refine the generated result.</p></left>-->
<!-- <br>-->
<!-- <center><img class="card-img-top" src="materials/framework.png" style="width:400px"></center>-->
<!-- <br><br>-->
<!-- <!– <video width="800" loop autoplay muted style="border:1px solid black"> –>-->
<!-- <video width="850" loop autoplay muted controls>-->
<!-- <source src="materials/new3-2.mp4" type="video/mp4">-->
<!-- </video>-->
<!-- </div>-->
<!-- <br><br>-->
<!-- <p><b>Image generation: </b> A pre-trained diffusion model is used as the generator, and multiple scorers, such as CLIP and image classifiers, are used to provide feedback to the generator.</p>-->
<!-- <p><b>Video question answering: </b> GPT-2 is used as the generator, and a set of CLIP models are used as scorers.</p>-->
<!-- <p><b>Grade school math: </b> GPT-2 is used as the generator, and a set of question-solution classifiers are used as scorers.</p>-->
<!-- <p><b>Robot manipulation: </b> MPC+World model is used as the generator, and a pre-trained image segmentation model is used to compute the scores from multiple camera views to select the best action.</p>-->
<!-- </div>-->
</section>
<section id="interp">
<hr>
<h2>Aligner Interpretability</h2>
<br><br>
<div style="display: flex; justify-content: space-between; align-items: center;">
<img src="images/intep_correction.png" style="width: 30%;">
<img src="images/intep_copy.png" style="width: 28.35%;">
<img src="images/intep_representation_control.png" style="width: 40%;">
</div>
<br>
<figcaption style="text-align: left;">
<strong>Interpretability of the <i>Aligner</i> module.</strong>
<br>
Interpretability experiment results on <i>Aligner</i>: <strong>(a)(b)</strong> The LAT scan graph of <i>Aligner</i>'s each layer when generating the first 20 output tokens for two given question-answer pairs. A higher value in the graph indicates a more active correction representation in that layer. Specifically, (a) exhibits raised activity, suggesting an enhanced correction action in the output, whereas (b) displays a tendency towards copying the original response. Moreover, the distinct differences between the two graphs are mainly observed in the early layers. This indicates that the decision regarding the degree of correction is made in the early layers of <i>Aligner</i>. <strong>(c)</strong> The control experiment shows the effectiveness of the extracted correction representation vector in modulating the <i>Aligner</i>'s correction behavior. The relationship between the Average Levenshtein Ratio and representation vector coefficients is approximately linear, with an \(R^2\) value of approximately 0.93.
</figcaption>
</section>
<section id="results">
<hr>
<h2>Results Overview</h2>
<br><br>
<figure class="mx-auto">
<center><img class="card-img-top" src="materials/4in1_1209.jpg" style="width:950px"></center>
<br>
<figcaption style="text-align: left;">
<strong>Distribution of helpfulness and harmlessness scores in training and evaluation sets. </strong>
<br>
<b>(a)</b> The distribution shift in answers and correctional answers in the training dataset; <br>
<b>(b)</b> redistribution shift of <i>Aligner</i>-7B, based on upstream models such as GPT-4 (b1), Alpaca-7B (b2) and Llama2-70B-Chat (b3).<br>
Based on the figure, we found that:<br>
<strong>(1)</strong> The correctional answer in the training dataset surpasses the original answers in terms of both helpfulness and harmlessness; <br>
<strong>(2)</strong> The refuse-to-answer pattern of GPT-4 created an area of overcorrected answers where both helpful and harmless scores are low, and our <i>Aligner</i>-7B improved these answers by providing additional information and corrections. <br>
<strong>(3)</strong> The Alpaca-7B model, which is not aligned, had its answers corrected by our <i>Aligner</i>-7B, significantly increasing both scores. <br>
<strong>(4)</strong> The Llama2-70B-Chat model is already aligned (the average safety score is higher than the correction in the training dataset), and the correction of <i>Aligner</i>-7B enhanced the helpfulness significantly while maintaining the harmless score.<br>
<strong>For detailed performance result of <i>Aligner</i> models, please refer to <a href="#detail">Detailed Results</a>.</strong>
</figcaption>
</figure>
<hr>
<br>
<!-- <div class="mx-auto" style="display: flex; justify-content: center; align-items: center;">
<img src="materials/test_gpt_4.gif" style="width: 450px; margin-right: 50px;">
<img src="materials/test_gpt_2.gif" style="width: 450px;">
</div> -->
<div class="mx-auto" style="display: flex; justify-content: center; align-items: center;">
<img src="materials/test_gpt_4.gif" style="width: 550px">
<center><figcaption style="text-align: left;">
<b>Distribution shift of helpful and harmless scores in the training process of <i>Aligner</i>-7B model</b><br>
This figure shows the distribution shift of helpfulness and harmlessness scores in the evaluation of checkpoint models of our <i>Aligner</i>-7B model. In the training process, the model has quickly learned the correction pattern in a relatively short time, and the learning process exhibits strong transparency and parameter efficiency.
</figcaption></center>
</div>
<br>
</section>
<section id="detail">
<hr>
<h2>Detailed Results</h2>
<br><br>
<!-- <figure class="mx-auto">
<center><img class="card-img-top" src="materials/table_2.jpg" style="width:750px"></center>
<br>
<figcaption style="text-align: left;">
<b><i>Weak-to-strong generalization</i> results</b><br>
The results demonstrate that <i>Aligner</i>-7B can achieve <i>weak-to-strong generalization</i> on 7B, 13B, and 70B upstream models with existing alignment methods using the labels given by the <i>Aligner</i>. This process entails enhancing the capabilities of a stronger model by finetuning it with labels generated from a weaker model.
</figcaption>
</figure>
<hr>
<br> -->
<figure class="mx-auto">
<center><img class="card-img-top" src="materials/Tab4_new.jpg" style="width:950px"></center>
<br>
<figcaption style="text-align: left;">
<b>Ablation Study: Different <i>Identity Mapping</i> Proportions</b><br>
We first trained an identity <i>Aligner</i> for identity mapping, followed by extensive residual Q-A-C learning based on this <i>Aligner</i>. Specifically, we formed the Q-A-A dataset by extracting partial data from the training dataset in proportions of 2%, 10%, 20%, and 50%. The table presents our control experiments with a 50K training dataset, showing that extracting 20% of the data (<i>i.e.</i>, 10K dataset size) for initial constant identity training yields relatively better results.
</figcaption>
</figure>
<hr>
<br>
<!-- <div class="mx-auto" style="display: flex; justify-content: center; align-items: center;">
<img src="materials/test_gpt_4.gif" style="width: 450px; margin-right: 50px;">
<img src="materials/test_gpt_2.gif" style="width: 450px;">
</div> -->
<!-- <figure class="mx-auto">
<center><img class="card-img-top" src="materials/table_6.jpg" style="width:950px"></center>
<br>
<figcaption style="text-align: left;">
<b>Performance of <i>Aligner</i> Models</b><br>
It is shown that <i>Aligner</i> achieves significant performances in all the settings. All assessments in this table were conducted based on integrating various models with <i>Aligners</i> to compare with the original models to quantify the percentage increase in helpfulness and harmlessness. The background color represents the type of target language model: green represents API-based models, orange represents open-source models without safety alignment, and blue represents safety-aligned open-source models. The icon <i class="fas fa-lock" style="color: rgba(0, 0, 0, 0.7);"></i> indicates the model parameters are not accessible and <i class="fas fa-shield-alt" style="color: rgba(0, 0, 0, 0.7);"></i> indicates the model is safety-aligned.
</figcaption>
</figure>
<hr>
<br> -->
<figure class="mx-auto">
<center><img class="card-img-top" src="materials/table_7.jpg" style="width:550px"></center>
<br>
<figcaption style="text-align: left;">
Comparison to Self-Refine, Critique Methods<br>
Constitutional AI (CAI) <cite>
Constitutional ai: Harmlessness from ai feedback.
</cite>, Self-Critique <cite>
Self-critiquing models for assisting human evaluators
</cite>, and Self-Refine <cite>
Self-Refine: Iterative Refinement with Self-Feedback
</cite>, primarily utilize the self-critiquing and refining capabilities of LLMs to enhance their performance. We employ CAI prompts solely during the inference time of LLMs to encourage self-revision of their answers. As demonstrated in the table above, our method, <i>Aligner</i>, outperforms the baseline considering both helpfulness and harmlessness dimensions. Additionally, baseline methods typically require multiple dialogue iterations and extended context windows for prompt insertion and ongoing self-correction. This could result in longer inference times and considerable consumption of context window length.
</figcaption>
</figure>
<hr>
<br>
<figure class="mx-auto">
<center><img class="card-img-top" src="materials/table_8.jpg" style="width:550px"></center>
<br>
<figcaption style="text-align: left;">
Performance of <i>Aligner</i> on the Various Preference Datasets<br>
To demonstrate the independence of <i>Aligner</i> from specific datasets, we utilized various open-source RLHF preference datasets. Specifically, we trained on HH-RLHF<cite>
Training a helpful and harmless assistant with reinforcement learning from human feedback.
</cite> and PKU-SafeRLHF <cite>
Beavertails: Towards improved safety alignment of LLM via a human-preference dataset.
</cite><cite>
Safe rlhf: Safe reinforcement learning from human feedback.
</cite> datasets and compared <i>Aligner</i> with SFT, RLHF, and DPO. After fine-tuning Alpaca-7B with SFT, RLHF, and DPO, we compare these models against the original Alpaca-7B corrected by <i>Aligner</i>. The experiment results (as shown in the table above) indicate that <i>Aligner</i>'s performance in enhancing the original model's capabilities is comparable to, or exceeds, that of the baseline methods. Notably, models finetuned with RLHF or DPO tend to generate either conservative answers or fail to recognize dangers while adding helpful information explicitly. Importantly, training with RLHF or DPO methods requires optimizing significantly more models and consuming more training resources than just training an <i>Aligner</i>, <i>e.g.</i>, for a 70B model, DPO needs 11.25 times and RLHF 22.5 times more resources than <i>Aligner</i>.
</figcaption>
</figure>
</section>
<section id="application">
<hr>
<h2>Applications</h2>
<p style="font-size: 23px;"><b>Multi-round RLHF training via Aligner</b></p>
<p style="font-size: 20px;"><b>New Multi-round Training Pipeline</b></p>
<p>As a data augmentation tool, <i>Aligner</i> can enhance the upstream model's response \(A\) into an improved response \(A^*\), thereby forming a synthetic preference dataset. This dataset can be used to further train the upstream model via RLHF/DPO. Repeating this process allows for multi-round RLHF or DPO.</p>
<figure class="mx-auto">
<center><img class="card-img-top" src="images/multi_round_pipeline.png" style="width:950px"></center>
<br>
<figcaption style="text-align: left;">
This paradigm brings many <b>advantages</b>:
<ul>
<li>The <i>Aligner</i> inherits the feature of transferring from the dispreferred distribution to the preferred distribution in the preference dataset.</li>
<li><i>Aligner</i> modifies the upstream model to produce better answers, bringing the distribution of resulting preference dataset closer to the answer distribution of the upstream model. This effectively mitigates the reward model collapse problem caused by out-of-distribution (OOD) preference datasets.</li>
<li>The <i>Aligner</i> serves as a synthetic data generator, providing an efficient and repeatable method for constructing preference datasets.</li>
</ul>
</figcaption>
</figure>
<p style="font-size: 20px;"><b>Performance</b></p>
<figure class="mx-auto">
<center><img class="card-img-top" src="images/multi_round_performance.png" style="width:750px"></center>
<br>
<figcaption style="text-align: left;">
We conducted three rounds of RLHF and DPO on Alpaca2-7B using the three-round preference dataset from PKU-SafeRLHF. Following this, we trained three rounds of <i>Aligner</i>s with the same three-round preference datasets, which were then employed to refine the upstream model and generate new preference datasets. These synthetic preference datasets were subsequently used to fine-tune the upstream model. As illustrated in Figure above, the introduction of <i>Aligner</i>-corrected new preference datasets enabled the upstream model to improve both utility and safety simultaneously. In contrast, a typical multi-round RLHF/DPO pipeline only enhances utility, leaving the responses unsafe.
</figcaption>
</figure>
<p style="font-size: 23px;"><strong>Weak-to-Strong Correction via Aligner</strong></p>
<p>As AI systems reach human-level performance across various tasks and undertake increasingly complex activities that are hard for humans to grasp, it becomes progressively challenging to provide ongoing, reliable feedback and ensure that their behaviors align with human intentions. This brings forth the significant issue of the Superalignment problem: <em>How can we deliver supervisory signals to advanced AI systems and ensure they remain aligned with human goals?</em><cite>
AI Alignment: a Comprehensive Survey