-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
339 lines (307 loc) · 20.1 KB
/
Copy pathindex.html
File metadata and controls
339 lines (307 loc) · 20.1 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
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
<!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="We study sequential LLM training (pretraining, post-training, fine-tuning) and show that mixing a small amount of target data into pretraining improves robustness of post-trained capabilities to forgetting under downstream fine-tuning, even when post-training metrics look unchanged.">
<meta property="og:title" content="Early Data Exposure Improves
Robustness to Subsequent Fine-Tuning"/>
<meta property="og:description" content="Early exposure to target data during pretraining improves retention after downstream fine-tuning while keeping post-training performance matched."/>
<meta property="og:url" content="https://grghosal.github.io/earlyexposurewebdraft/"/>
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="static/images/8panel.png" />
<meta property="og:image:width" content="1200"/>
<meta property="og:image:height" content="630"/>
<meta name="twitter:title" content="Early Data Exposure Improves Robustness to Subsequent Fine-Tuning">
<meta name="twitter:description" content="Early exposure to target data during pretraining improves retention after downstream fine-tuning while keeping post-training performance matched.">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="static/images/8panel.png">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="Large language models, fine-tuning, forgetting, sequential training, early exposure, pretraining, LLM">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Early Data Exposure Improves Robustness to Subsequent Fine-Tuning </title>
<link rel="icon" type="image/x-icon" href="static/images/8panel.png">
<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?v=9">
<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">Early Data Exposure Improves Robustness to Subsequent Fine-Tuning</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block"><a href="https://lawrencefeng17.github.io/" target="_blank">Lawrence Feng</a></span>,
<span class="author-block">
<a href="https://grghosal.github.io/" target="_blank">Gaurav R. Ghosal</a></span>,
<span class="author-block">Jacob Mitchell Springer</span>,
<span class="author-block">Ziqian Zhong</span>,
<span class="author-block">
<a href="https://www.cs.cmu.edu/~aditirag/" target="_blank">Aditi Raghunathan</a>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">Carnegie Mellon University</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/2507.09937" 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>
<!-- GitHub link -->
<span class="link-block">
<a href="https://github.com/grghosal/earlyexposurewebdraft" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>GitHub</span>
</a>
</span>
<!-- ArXiv abstract Link -->
<span class="link-block">
<a href="https://arxiv.org/abs/2507.09937" 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">
<center>
<img src="static/images/8panel.png" alt="Figure 1" style="width: 100%; max-width: 100%; height: auto; display: block; margin: 10px auto;">
</center>
<h2 class="subtitle has-text-centered figure-caption">
Standard training of LLMs acquires capabilities during post-training that are often lost during later fine-tuning. As a result, models struggle to retain important behaviors when adapted to new tasks. We find that mixing a small amount of target data into pretraining improves robustness to forgetting, enabling better retention after downstream training.
</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">TL;DR</h2>
<div class="content has-text-justified">
<ul style="font-size: 1.2em; line-height: 1.6;">
<li><strong style="color: #b91c1c;">The Problem:</strong> Post-trained capabilities are fragile — fine-tuning on new tasks can rapidly erase previously learned capabilities.</li>
<li><strong style="color: #2563eb;">The Insight:</strong> When a model learns matters as much as what it learns. Introducing target data early (during pretraining) changes how the capability is represented — shifting it into more isolated, robust features that resist later overwriting.</li>
<li><strong style="color: #15803d;">The Result:</strong> Early exposure improves retention without hurting performance — shifting the tradeoff frontier and making capabilities resistant to forgetting.</li>
</ul>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- Paper poster -->
<section class="hero is-small is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title">The Retention Problem in LLM Pipelines</h2>
<div class="content has-text-justified">
<img class="content-figure" src="static/images/RetentionAndMultiStage.png" alt="Multi-stage training: acquiring capability X and retention after fine-tuning for task Y" style="width: 100%; max-width: 100%; height: auto; display: block; margin: 10px 0;">
<p class="figure-caption">
<b><i>Figure 2: Sequential training pipeline in LLMs.</i></b> A model is first pretrained (θ<sub>pre</sub>) to learn general representations, then post-trained to acquire a specific capability <i>X</i> (θ<sub>post</sub>), and finally fine-tuned by a downstream user for a new task <i>Y</i> (θ<sub>ft</sub>). The key question is whether the capability learned during post-training is preserved after downstream fine-tuning, highlighting the potential for later updates to modify or overwrite earlier knowledge.
</p>
<p>
Modern training pipelines for large language models proceed through multiple sequential stages (e.g., pretraining → post-training → fine-tuning), each updating the same underlying parameters. This implicitly assumes that newly learned capabilities can be added without disrupting existing ones. We study this in two stages:
</p>
<div class="experiment-details">
<ul style="margin-left: 20px;">
<li><strong>Upstream acquisition:</strong> Learning a target capability <i>X</i> during post-training</li>
<li><strong>Downstream adaptation:</strong> Fine-tuning the model for a new task <i>Y</i></li>
</ul>
</div>
<div style="background-color: rgba(185, 28, 28, 0.12); padding: 16px 20px; margin: 10px 0; width: 100%; border-radius: 0; box-sizing: border-box; border-left: 4px solid rgba(185, 28, 28, 0.65);">
<p style="margin: 0 0 10px 0; font-size: 1.45em; font-weight: bold; color: #b91c1c;">Key Challenge</p>
<p style="margin: 0; font-weight: bold;">
Capabilities acquired earlier in the pipeline are not inherently preserved—subsequent training can overwrite or degrade them.
</p>
</div>
</div>
</div>
</div>
</section>
<!--End paper poster -->
<section class="hero is-small is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title">Immediate Performance ≠ Long-term Retention</h2>
<div class="content has-text-justified">
<img class="content-figure" src="static/images/SamePTDifferentRetention.png" alt="Illustration: identical immediate post-training performance with differing retention after downstream fine-tuning" style="width: 100%; max-width: 100%; height: auto; display: block; margin: 10px 0;">
<p class="figure-caption">
<b><i>Figure 3: Same post-training performance, differing retention.</i></b> Mixing doesn’t help immediately—but it matters later. As λ (fraction of data mixed during pre-training) increases, post-training loss stays flat, while retention after fine-tuning improves.
</p>
<p>
We study how early exposure to post-training data affects downstream retention. We vary how much of the target data is introduced during pretraining (λ), then train all models to convergence on the same post-training objective before fine-tuning on a new task.
</p>
<div class="experiment-details">
<ul style="margin-left: 20px;">
<li><strong>Controlled setup:</strong> Vary λ while keeping post-training procedure fixed</li>
<li><strong>Equalized performance:</strong> All models reach similar loss after post-training</li>
<li><strong>Downstream evaluation:</strong> Fine-tune on a new task to induce forgetting</li>
</ul>
</div>
<p>
We find that across λ (the ratio of post-training data mixed in during pre-training), models reach nearly the same MusicPile loss after post-training, demonstrating that early exposure has a negligible impact on immediate capability learning in our setting. After downstream fine-tuning, however, models with more early exposure forget less and retain substantially lower MusicPile loss. Our results illustrate that two models can look equally capable immediately after post-training, yet differ sharply in how much of that capability survives later updates.
</p>
<div style="background-color: rgba(37, 99, 235, 0.12); padding: 16px 20px; margin: 10px 0; width: 100%; border-radius: 0; box-sizing: border-box; border-left: 4px solid rgba(37, 99, 235, 0.65);">
<p style="margin: 0 0 10px 0; font-size: 1.45em; font-weight: bold; color: #2563eb;">Takeaway</p>
<p style="margin: 0; font-weight: bold;">
Immediate performance is unchanged across λ, but retention improves with early exposure—models that see more data during pretraining consistently forget less after fine-tuning.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="hero is-small is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title">Where You Put the Data Matters</h2>
<div class="content has-text-justified">
<img class="content-figure" src="static/images/DataPlacement.png" alt="Schematic of how target data placement across pretraining versus post-training affects representations" style="width: 100%; max-width: 100%; height: auto; display: block; margin: 10px 0;">
<p class="figure-caption">
<b>Figure 4:</b> Under a fixed budget of steps on post-training data, increasing the fraction of data used during pretraining (λ) worsens immediate post-training loss but improves retained performance after fine-tuning. This highlights a tradeoff between short-term optimization and long-term retention.
</p>
<p>
Early exposure helps retention — but is that because the model sees the post-training data earlier, or simply because it sees more of it? To separate these effects, we fix the total amount of post-training data the model sees and vary only where that data is placed in the training pipeline.
</p>
<p>
We compare different allocations:
</p>
<div class="experiment-details">
<ul style="margin-left: 20px;">
<li><strong>All post-training:</strong> use the full data budget during the post-training stage.</li>
<li><strong>Mixed allocation:</strong> expose the model to some of the data during pretraining, then reserve the rest for post-training.</li>
<li><strong>All early exposure:</strong> use the full data budget during pretraining.</li>
</ul>
</div>
<p>
Our results reveal a clear tradeoff: reserving more data for post-training improves immediate target-domain performance, since Stage 2 directly optimizes for that capability. But after downstream fine-tuning, the pattern reverses: models exposed to some of the data earlier retain the capability better, despite receiving less dedicated post-training exposure. Ultimately, post-training helps acquire the capability quickly, while early exposure helps it survive later updates—so the best allocation lies between the extremes of relying on pre-training or post-training solely.
</p>
<div style="background-color: rgba(37, 99, 235, 0.12); padding: 16px 20px; margin: 10px 0; width: 100%; border-radius: 0; box-sizing: border-box; border-left: 4px solid rgba(37, 99, 235, 0.65);">
<p style="margin: 0 0 10px 0; font-size: 1.45em; font-weight: bold; color: #2563eb;">
Data placement changes durability.
</p>
<p style="margin: 0; font-weight: bold;">
Even under a fixed data budget, allocating some target-domain data to pretraining improves retention after downstream fine-tuning, while reserving all of it for post-training gives the best immediate performance.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="hero is-small is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title">Post-Training Interventions for Retention</h2>
<div class="content has-text-justified">
<img class="content-figure" src="static/images/Post-TrainingInterventions.png" alt="Overview of post-training interventions and their effect on retention after fine-tuning" style="width: 100%; max-width: 100%; height: auto; display: block; margin: 10px 0;">
<p class="figure-caption">
<b>Figure 5:</b> Placeholder caption — summarize what the panels show (e.g., intervention types, metrics, and retention relative to a baseline).
</p>
<p>
Early exposure is not the only way to make post-trained capabilities more durable. If <em>when</em> a model first sees the data matters, then a natural follow-up is whether we can also improve retention by changing how the model is post-trained.
</p>
<p>
We study two simple interventions during post-training:
</p>
<div class="experiment-details">
<ul style="margin-left: 20px;">
<li><strong>Replay:</strong> mix a small amount of general pretraining data back into the post-training stage.</li>
<li><strong>Dropout:</strong> regularize the post-training update so the model does not rely too heavily on brittle representations.</li>
</ul>
</div>
<p>
Both methods are applied before the downstream fine-tuning stage. This lets us ask whether they merely improve the post-trained checkpoint immediately, or whether they actually help the capability survive later updates.
</p>
<p>
Across our experiments, replay and dropout improve the retention–adaptation frontier. Models trained with these interventions retain more of the post-trained capability after downstream fine-tuning, while still adapting well to the new task. However, these gains do not replace early exposure. The strongest frontiers come from combining pretraining-time mixing with post-training interventions, suggesting that they operate through complementary mechanisms.
</p>
<div style="background-color: rgba(37, 99, 235, 0.12); padding: 16px 20px; margin: 10px 0; width: 100%; border-radius: 0; box-sizing: border-box; border-left: 4px solid rgba(37, 99, 235, 0.65);">
<p style="margin: 0 0 10px 0; font-size: 1.45em; font-weight: bold; color: #2563eb;">
Retention is shaped across the whole upstream pipeline.
</p>
<p style="margin: 0; font-weight: bold;">
Early exposure makes capabilities less brittle before post-training begins, while replay and dropout further improve durability during post-training itself.
</p>
</div>
</div>
</div>
</div>
</section>
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@misc{feng2025early,
title={Early Data Exposure Improves Robustness to Subsequent Fine-Tuning},
author={Lawrence Feng and Gaurav R. Ghosal and Jacob Mitchell Springer and Ziqian Zhong and Aditi Raghunathan},
year={2025},
eprint={2507.09937},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2507.09937}
}</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>