-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
457 lines (385 loc) · 18.6 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Progress or Regress? Self-Improvement Reversal in Post-training">
<meta name="keywords" content="LLMs, Self-Improvement, Post-training">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Progress or Regress? Self-Improvement Reversal in Post-training</title>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script>
<script type="text/javascript" src="static/js/sort-table.js" defer></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="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/trust.png">
<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>
<script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
</head>
<body>
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<div class="navbar-item is-hoverable">
<a class="navbar-item" href="#overview">
Overview
</a>
<a class="navbar-item" href="#systemformulation">
Systematic Formulation
</a>
<a class="navbar-item" href="#evaluation">
Reversal Phenomenon
</a>
<a class="navbar-item" href="#contact">
Contact
</a>
<a class="navbar-item" href="#reference">
BibTeX
</a>
</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">Progress or Regress? Self-Improvement Reversal in Post-training</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">Ting Wu<sup>1,3</sup> </span>
<span class="author-block">Xuefeng Li<sup>2,3</sup> </span>
<span class="author-block">Pengfei Liu<sup>2,3,4</sup> </span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Fudan University</span>
<span class="author-block"><sup>2</sup>Shanghai Jiao Tong University</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>3</sup>Generative AI Research Lab (GAIR)</span>
<span class="author-block"><sup>4</sup>Shanghai AI Laboratory</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/pdf/2407.05013" 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>
<span class="link-block">
<a href="https://github.com/GAIR-NLP/self_improvement_reversal" 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>
<!-- Dataset Link. -->
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<h2 class="subtitle has-text-centered">
<span style="font-weight: bold; color: #26a3e6"> Self-improvement Reversal </span> reveals the phenomenon that during iterative post-training, as pass@1 accuracy improves, broader capabilities like output diversity and out-of-distribution generalization decline, exhibiting a paradoxical trend.
</h2>
<div style="text-align: center;">
<img src="static/images/figure_case_study.png" width="80%" />
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3", id="overview">Overview</h2>
<div class="content has-text-justified">
<p>
🤔 <span style="font-weight: bold; font-size: 1.05em"> A thought-provoking research question </span>
</p>
<p>
Self-improvement through post-training methods has been acclaimed for enhancing the problem-solving capabilities (e.g., mathematical reasoning) of Large Language Models (LLMs) without human intervention. However, current research all concentrate on maximizing benchmark scores through iterative self-improvement, there is little
exploration of the underlying factors contributing to performance gains. As a result, the progress and reliability of different self-improvement methods are not guaranteed. Amidst the quest for self-improvement in LLMs, the persistent question arises: <b>are these iterative post-training methods truly fostering progress, or are they inadvertently leading to regression?</b>
</p>
<p>
👩💻 <span style="font-weight: bold; font-size: 1.05em"> Our research route </span>
</p>
<p>We first provide a comprehensive overview of the main iterative post-training paradigms for self-improvement, understanding both the explicit and implicit influencing factors that contribute to the consistent performance improvements. This provide actionable insights for practitioners on how to perform iterative self-improvement more effectively. </p>
<p>We further develop an evaluative framework equipped with a comprehensive suite of metrics to assess improvement problems, solutions diversity, and OOD capabilities within the iterative process, enabling us to scrutinize the actual improvements beneath self-improvement. </p>
<p>
🧐 <span style="font-weight: bold; font-size: 1.05em"> What we reveal? </span>
</p>
<ul>
<li><b>Answer Selection Optimization: </b>Iterative self-improvement hardly entails the acquisition of new problem-solving abilities, but rather the enhancement of the model’s correct answer selection within its generation space.</li>
<li><b>Trade-off with Output Diversity: </b>There exists a critical trade-off in iterative self-improvement: while aiming for higher accuracy, the diversity of outputs, which can be crucial for creativity and robustness in problem-solving, is compromised. </li>
<li><b>Capabilities Collapse: </b>Iterative post-training methods can exacerbate the generalization disparities across groups, inadvertently causing models to focus on easier problems rather than enhancing their ability to solve more complex ones. </li>
</ul>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<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", id="systemformulation">Systematic Formulation of Self-Improvement in Post-training</h2>
<div style="text-align: left;">
<p>We identify key variables that influence the optimal improving performance and trend during the iterative post-training process: foundation model M, problem-solving task D, iteration steps T and post-training method F. For our experimental setup, we choose M = {LLaMA2-7B, LLaMA3-8B, Mistral-7B}, D = {CommonsenseQA for Commonse Knowledge, GSM8K and MATH for Mathematical Reasoning, MBPP for Code Generation}, T = {1, 2, 3, 4, 5}, and F = {Iterative SFT, Iterative DPO, Iterative SFT-DPO}. </p>
<p> <br> </p>
</div>
<div align="center">
<figure>
<img src="static/images/result.png" width=90%/>
</figure>
</div>
<div style="text-align: left;">
<span style="font-weight: bold; font-size: 1.05em"> Explicit influencing Factors </span> Across all methods and datasets, there is a general trend of improvement in pass@1 accuracy with increasing iteration steps. This indicates that iterative post-training effectively enhances model performance over time. However, the rate of improvement tends to plateau or even decline slightly after 4-5 iterations.
</div>
<div align="center">
<figure>
<img src="static/images/coverage.png" width=100%/>
</figure>
</div>
<div style="text-align: left;">
<span style="font-weight: bold; font-size: 1.05em"> Implicit Underlying Factor </span> We found that <b>Correct Answer Coverage</b>, the proportion of the correct answer space that M<sub>1</sub> occupy, can be used to gauge the model’s gains in self-improvement. The table above clearly demonstrates that when the correct answer coverage is high (>0.5), Iterative DPO and Iterative SFT-DPO produce the best-performing M<sub>t</sub><sup>*</sup>. Conversely, when the coverage is lower (< 0.5), Iterative SFT is more effective in achieving the optimal M<sub>t</sub><sup>*</sup>. Therefore, correct answer coverage can serve as a key factor in guiding practitioners to choose the suitable post-training method F.
</div>
</div>
</div>
</div>
</section>
<section class="section" id="evaluation">
<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">Critical Evaluations on Self-Improvement</h2>
<div class="content has-text-justified">
<p>
We engage in a critical examination and reevaluation of iterative self-improvement: discerning whether the improvements constitute genuine progress or merely regression.
<h4 class="title is-8">🌟 Improvement Problems</h4>
<div align="center">
<figure>
<img src="static/images/improvementset.png" width=80%/>
</figure>
</div>
<p><span style="font-weight: bold"> Reversal Observation </span> As N grows, M1 achieves near-perfect pass@N accuracy on IS(t) (Improvement set), suggesting its inherent capacity to tackle the deemed improvement problems. </p>
<h4 class="title is-8">🌟 Solutions Diversity</h4>
<div align="center">
<figure>
<img src="static/images/solution.png" width=80%/>
</figure>
</div>
<p><span style="font-weight: bold"> Reversal Observation </span> All methods show a consistent decrease in diversity, significantly diminishing the diversity of model outputs over iterations</href>, impacting both correct and incorrect answers. This reduction is evident across all three metrics: syntactic (Distinct-N grams), semantic (SentenceBERT Consine Similarity), and logical (Distinct Equations) diversity.</p>
<h4 class="title is-8">🌟 OOD Generalization</h4>
<div align="center">
<figure>
<img src="static/images/ood.png" width=80%/>
</figure>
</div>
<p><span style="font-weight: bold"> Reversal Observation </span> With an increase in iterative steps, Iterative SFT and Iterative SFT-DPO can significantly impair OOD generalization. </href>In contrast, Iterative DPO shows a noticeable improvement. However, all three iterative post-training methods can exacerbate generalization disparities across groups, inadvertently causing models to focus on easier problems instead of improving their ability to solve more complex ones.</p>
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Contact. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3", id="contact">📬 Contact</h2>
<div class="content has-text-justified">
<p>
If you have any questions regarding this project, feel free to submit a <a href="https://github.com/GAIR-NLP/self-improvement-reversal">github issue</a> or reach out to us via email.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-four-fifths">
<h2 class="title", id="reference">BibTeX</h2>
<p>If you find our paper and code helpful, please consider citing our work😊
</p>
<pre><code>@artical{wu2024progressregressselfimprovementreversal,
title={Progress or Regress? Self-Improvement Reversal in Post-training},
author={Ting Wu and Xuefeng Li and Pengfei Liu},
year={2024},
eprint={2407.05013},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.05013}
}
</code></pre>
</div>
</div>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is 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>
This means you are free to borrow the <a
href="https://github.com/nerfies/nerfies.github.io">source code</a> of this website,
we just ask that you link back to this page in the footer.
Please remember to remove the analytics code included in the header of the website which
you do not want on your website.
</div>
</div>
</div>
</div>
</footer>
<script>
function showLeaderboard(type) {
// Hide all leaderboards initially
var allTables = document.querySelectorAll('.leaderboard-table');
allTables.forEach(table => table.style.display = 'none');
// Show and sort the selected leaderboard
var table, isAscending;
switch (type) {
case 'selfKnowledge':
table = document.getElementById('selfKnowledgeTable');
isAscending = false; // Higher values at the top
break;
case 'nonDeceptiveness':
table = document.getElementById('nonDeceptivenessTable');
isAscending = true; // Lower values at the top
break;
case 'consistency':
table = document.getElementById('consistencyTable');
isAscending = false; // Higher values at the top
break;
}
table.style.display = 'block';
sortTableByColumn(table.id, 1, isAscending);
}
function sortTableByColumn(tableId, colIndex, ascending = true) {
var table = document.getElementById(tableId);
var rows = Array.from(table.rows).slice(1);
rows.sort(function(a, b) {
var cellA = parseFloat(a.cells[colIndex].innerText) || 0;
var cellB = parseFloat(b.cells[colIndex].innerText) || 0;
return ascending ? cellA - cellB : cellB - cellA;
});
rows.forEach(function(row) {
table.tBodies[0].appendChild(row);
});
}
// Initialize the sorting for each leaderboard upon page load
document.addEventListener('DOMContentLoaded', function() {
showLeaderboard('selfKnowledge'); // Show and sort self-knowledge by default
});
document.addEventListener('DOMContentLoaded', function() {
document.querySelectorAll('.leaderboard-table th').forEach(function(header, index) {
header.addEventListener('click', function() {
var tableId = header.closest('table').id;
var ascending = !header.classList.contains('ascending');
sortTableByColumn(tableId, index, ascending);
header.classList.toggle('ascending', ascending);
});
});
});
</script>
<style>
.hidden {
display: none;
}
.sortable:hover {
cursor: pointer;
}
.asc::after {
content: ' ↑';
}
.desc::after {
content: ' ↓';
}
.button-64 {
align-items: center;
background-image: linear-gradient(144deg, #C972FF, #7D74F3 50%, #40E0EB); /* Lighter gradient colors */
border: 0;
border-radius: 8px;
box-shadow: rgba(151, 65, 252, 0.2) 0 15px 30px -5px;
box-sizing: border-box;
color: #FFFFFF;
display: flex;
font-family: Phantomsans, sans-serif;
font-size: 9px; /* Smaller font size */
justify-content: center;
line-height: 1em;
min-width: 70px; /* Smaller width */
padding: 2px; /* Less padding */
text-decoration: none;
user-select: none;
-webkit-user-select: none;
touch-action: manipulation;
white-space: nowrap;
cursor: pointer;
margin-right: 10px; /* Space between buttons */
}
.button-64:active,
.button-64:hover {
outline: 0;
}
.button-64 span {
background-color: rgb(235, 235, 245); /* Lighter background color for the text span */
color: #333; /* Darker text color for better contrast */
padding: 8px 12px; /* Adjust padding for smaller size */
border-radius: 6px;
width: 100%;
height: 100%;
transition: 300ms;
}
.button-64:hover span {
background: rgba(235, 235, 245, 0.5); /* Even lighter on hover */
}
@media (min-width: 768px) {
.button-64 {
font-size: 16px; /* Adjust font size for larger screens */
min-width: 120px; /* Adjust minimum width for larger screens */
}
}
</style>
</body>
</html>