-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
261 lines (248 loc) · 9.94 KB
/
index.html
File metadata and controls
261 lines (248 loc) · 9.94 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
<!doctype html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>DataProphet | Demystifying Supervision Data Generalization in MLLMs</title>
<meta
name="description"
content="Project page for DataProphet: a training-free metric for predicting supervision data influence in multimodal LLMs."
/>
<link rel="preconnect" href="https://fonts.googleapis.com" />
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
<link
href="https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@400;500;700&family=Source+Serif+4:opsz,[email protected],400;8..60,600&display=swap"
rel="stylesheet"
/>
<link rel="stylesheet" href="./styles.css" />
</head>
<body>
<div class="bg-orb bg-orb-a"></div>
<div class="bg-orb bg-orb-b"></div>
<header class="site-header">
<a class="brand" href="#top">
<img src="./assets/images/icon.png" alt="DataProphet icon" />
<span>DataProphet</span>
</a>
<nav class="nav-links">
<a href="#abstract">Abstract</a>
<a href="#analysis">Analysis</a>
<a href="#metric">Metric</a>
<a href="#selection">Selection</a>
<a href="#citation">Citation</a>
</nav>
</header>
<main id="top">
<section class="hero reveal">
<div class="hero-copy">
<p class="tagline">ICLR 2026 Submission</p>
<h1>DataProphet: Demystifying Supervision Data Generalization in Multimodal LLMs</h1>
<p class="authors">Xuan Qi, Luxi He, Dan Roth, Xingyu Fu</p>
<p class="hero-summary">
DataProphet predicts which supervision datasets will help a target benchmark before any
training. It combines multimodal similarity, perplexity, and diversity into a
training-free transfer score.
</p>
<div class="hero-buttons">
<a class="btn btn-primary" href="./assets/paper/dataprophet-paper.pdf" target="_blank" rel="noreferrer">Download PDF</a>
<a class="btn btn-secondary" href="https://huggingface.co/datasets/THUQiXuan/DataProphet" target="_blank" rel="noreferrer">Hugging Face</a>
<a class="btn btn-secondary" href="https://github.com/DataProphet26/dataprophet" target="_blank" rel="noreferrer">GitHub Code</a>
</div>
<div class="stats-grid">
<article>
<h3>14</h3>
<p>Source/target datasets</p>
</article>
<article>
<h3>7</h3>
<p>Task families</p>
</article>
<article>
<h3>0.860</h3>
<p>Kendall's tau (avg)</p>
</article>
<article>
<h3>+6.9%</h3>
<p>Synthetic data gain</p>
</article>
</div>
</div>
<div class="hero-visual">
<img src="./assets/images/icon2.png" alt="DataProphet illustration" />
</div>
</section>
<section id="abstract" class="panel reveal">
<h2>Abstract</h2>
<p>
Conventional data selection for multimodal LLMs often follows intuitive task similarity,
but this paper shows that intuition is unreliable for predicting transfer gains. The
authors evaluate 14 vision-language datasets across 7 task families and find that
influence is asymmetric and dataset-specific. DataProphet introduces a simple training-free
metric that integrates question/answer/image similarity, multimodal perplexity, and source
diversity to predict transfer ranking. The predicted rankings strongly correlate with actual
fine-tuning outcomes (Kendall's tau 0.860), and DataProphet-guided selection improves
average performance over uniform and training-based baselines under fixed compute budgets.
</p>
</section>
<section class="panel reveal">
<h2>Core Takeaways</h2>
<div class="takeaway-grid">
<article>
<h3>Task similarity is a weak proxy</h3>
<p>
OCR supervision can improve spatial reasoning more than chart tasks. Transfer cannot be
inferred from high-level task labels alone.
</p>
</article>
<article>
<h3>Influence is directional</h3>
<p>
The gain from train -> test is not symmetric: Delta<sub>s->t</sub> and
Delta<sub>t->s</sub> can differ substantially.
</p>
</article>
<article>
<h3>Training-free selection can win</h3>
<p>
DataProphet reaches +3.4% average improvement on real-data reweighting and +6.9% on
synthetic data selection versus uniform sampling.
</p>
</article>
</div>
</section>
<section id="analysis" class="panel reveal">
<h2>Data Influence Analysis</h2>
<p class="panel-lead">
Controlled fine-tuning with InternVL3-2B on each source dataset (20K samples) reveals
non-intuitive cross-task transfer patterns.
</p>
<figure>
<img src="./assets/images/teaser.png" alt="DataProphet teaser figure with three key findings" />
<figcaption>Figure: Three major takeaways from the paper.</figcaption>
</figure>
<figure>
<img src="./assets/images/relative.png" alt="Relative improvement heatmap across 14 train and test datasets" />
<figcaption>
Relative improvement heatmap (train dataset on y-axis, test dataset on x-axis).
</figcaption>
</figure>
</section>
<section id="metric" class="panel reveal">
<h2>The DataProphet Metric</h2>
<p class="equation">
M(s->t) = (QSim * ASim * ISim * PPL(s) * (Sil + H)) / PPL(t)
</p>
<p>
The metric is directional and training-free. It rewards source datasets that are aligned
with the target in text and vision space, challenging enough to teach new capability, and
diverse in question coverage.
</p>
<div class="metric-grid">
<figure>
<img src="./assets/images/metric.png" alt="DataProphet metric heatmap values" />
<figcaption>Metric score matrix between all source-target pairs.</figcaption>
</figure>
<div class="metric-notes">
<h3>Ranking Quality</h3>
<ul>
<li>Average tau<sub>Tgt</sub> = 0.863</li>
<li>Average tau<sub>Src</sub> = 0.857</li>
<li>Overall average tau = 0.860</li>
</ul>
<h3>Ablation Signal</h3>
<ul>
<li>Removing perplexity: 0.860 -> 0.487</li>
<li>Removing image similarity: 0.860 -> 0.625</li>
<li>Removing diversity: 0.860 -> 0.659</li>
</ul>
</div>
</div>
<figure>
<img src="./assets/images/result-kendall.png" alt="Kendall tau results by dataset" />
<figcaption>Kendall tau across target datasets.</figcaption>
</figure>
</section>
<section id="selection" class="panel reveal">
<h2>Data Selection Results</h2>
<p class="panel-lead">
Under a fixed budget of 280K samples, DataProphet-guided selection outperforms both uniform
and training-based methods in real and synthetic settings.
</p>
<div class="table-wrap">
<table>
<thead>
<tr>
<th>Setting</th>
<th>Uniform</th>
<th>ICONS</th>
<th>Oracle</th>
<th>DataProphet</th>
</tr>
</thead>
<tbody>
<tr>
<td>Real Data Reweighting (Avg)</td>
<td>67.6</td>
<td>69.6</td>
<td>70.8</td>
<td><strong>71.0</strong></td>
</tr>
<tr>
<td>Improve over Uniform</td>
<td>-</td>
<td>+2.0</td>
<td>+3.2</td>
<td><strong>+3.4</strong></td>
</tr>
<tr>
<td>Synthetic Data Selection (Avg)</td>
<td>55.1</td>
<td>60.8</td>
<td>-</td>
<td><strong>62.0</strong></td>
</tr>
<tr>
<td>Improve over Uniform</td>
<td>-</td>
<td>+5.7</td>
<td>-</td>
<td><strong>+6.9</strong></td>
</tr>
</tbody>
</table>
</div>
<div class="selection-notes">
<article>
<h3>Synthetic Source Mix</h3>
<p>
Among selected synthetic samples, approximately 38% come from GPT-5 and 62% from
Gemini 2.5 Pro.
</p>
</article>
<article>
<h3>RL Post-training</h3>
<p>
DataProphet allocation improves average score from 0.583 -> 0.595 (real RL data) and
0.564 -> 0.577 (synthetic RL data).
</p>
</article>
</div>
</section>
<section id="citation" class="panel reveal">
<h2>Citation</h2>
<p>If this work is useful, please cite:</p>
<pre id="bibtex"><code>@article{qi2026dataprophet,
title={DataProphet: Demystifying Supervision Data Generalization in Multimodal LLMs},
author={Qi, Xuan and He, Luxi and Roth, Dan and Fu, Xingyu},
journal={International Conference on Learning Representations},
year={2026}
}</code></pre>
<button class="btn btn-primary" id="copyBibBtn" type="button">Copy BibTeX</button>
</section>
</main>
<footer class="site-footer">
<p>DataProphet Project Page | Built from paper assets and LaTeX source</p>
</footer>
<script src="./script.js"></script>
</body>
</html>