-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
409 lines (365 loc) · 18.8 KB
/
Copy pathindex.html
File metadata and controls
409 lines (365 loc) · 18.8 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
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Taming Hallucinations: Boosting MLLMs’ Video Understanding via Counterfactual Video Generation">
<meta name="keywords" content="Taming Hallucinations, Boosting MLLMs’ Video Understanding, Counterfactual Video Generation">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Taming Hallucinations</title>
<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/favicon.svg">
<script src="https://cdn.jsdelivr.net/npm/chart.js"></script>
<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>
<style>
.container { max-width: 1100px; margin: 0 auto; }
/* Dataset stats tweaks */
.stats-grid { align-items: stretch; gap: 1rem; margin-left: 0 !important; margin-right: 0 !important; }
.stats-grid .column { display: flex; padding-left: 0; padding-right: 0; }
.stats-grid .box { padding: 1rem; display: flex; flex-direction: column; justify-content: center; height: 100%; width: 100%; }
.stats-grid .title { margin-bottom: 0.5rem; }
.stats-grid .table { font-size: 0.85rem; }
.stats-grid .table-container { flex: 0 0 auto; max-width: 100%; overflow-x: auto; }
/* Avoid overflow on desktop: subtract gaps from column width */
@media screen and (min-width: 1024px) {
.stats-grid .column.is-4-desktop {
flex: 0 0 calc((100% - 2rem) / 3);
max-width: calc((100% - 2rem) / 3);
}
}
/* LLM QA example image */
.qa-image { display: block; margin: 0 auto; width: 100%; max-width: 900px; height: auto; }
/* Center example video */
.video-center { display: block; margin: 0 auto; max-width: 100%; }
</style>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h3 class="title is-1 publication-title">Taming Hallucinations: Boosting MLLMs’ Video Understanding via Counterfactual Video Generation</h3>
<!-- <h2 class="title is-3 publication-title" style="color:#008bf5;">ICLR 2026</h2> -->
<div class="is-size-5 publication-authors">
<a href="https://github.com/Hertz19" target="_blank">Zhe Huang</a><sup>1,3*,†</sup>,
<a href="https://github.com/Costwen" target="_blank">Hao Wen</a><sup>2,3*,†</sup>,
<a href="https://scholar.google.com/citations?user=E7ey4lwAAAAJ&hl=en" target="_blank">Aiming Hao</a><sup>3*</sup>,
<a href="https://github.com/songbingze" target="_blank">Bingze Song</a><sup>3</sup>,
</div>
<div class="is-size-5 publication-authors">
<a href="https://scholar.google.com/citations?hl=en&user=fGc7NVAAAAAJ" target="_blank">Meiqi Wu</a><sup>3</sup>,
<a href="https://scholar.google.com/citations?user=jn21pUsAAAAJ&hl=en&oi=ao" target="_blank">Xiangxiang Chu</a><sup>3</sup>,
<a href="https://scholar.google.com/citations?hl=en&user=zCljkooAAAAJ" target="_blank">Jiahong Wu</a><sup>3‡</sup>,
<a href="https://scholar.google.com/citations?user=_8lB7xcAAAAJ" target="_blank">Sheng Lu</a><sup>2✉</sup>,
<a href="https://scholar.google.com/citations?user=eldgnIYAAAAJ&hl=zh-CN" target="_blank">Haoqian Wang</a><sup>1✉</sup>
</div>
<div class="is-size-5 publication-authors">
<sup>1</sup>Tsinghua University
<sup>2</sup>Beihang University
<sup>3</sup>AMAP, Alibaba Group
</div>
<div class="is-size-5 publication-authors">
<sup>*</sup>Co-first Authors
<sup>†</sup>Internship
<sup>‡</sup>Project Leader
<sup>✉</sup>Corresponding Author
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<!-- <span class="link-block">
<a href="" 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://arxiv.org/pdf/2512.24271" 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>
<!-- Video Link. -->
<!-- <span class="link-block">
<a href="https://www.youtube.com/watch?v=MrKrnHhk8IA"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-youtube"></i>
</span>
<span>Video</span>
</a>
</span> -->
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/AMAP-ML/Taming-Hallucinations" 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 -->
<span class="link-block">
<a href="https://huggingface.co/datasets/GD-ML/Taming-Hallucinations" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fa fa-database"></i>
</span>
<span>Dataset</span>
</a>
</span>
<!-- Demo Link. -->
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="" style="padding-bottom: 3rem">
<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">Abstract</h2>
<div class="content has-text-justified">
<p>
Multimodal Large Language Models (MLLMs) have made remarkable progress in video understanding. However, they suffer from a critical vulnerability: an over-reliance on language priors, which can lead to "<em>visual ungrounded hallucination</em>", especially when processing counterfactual videos that defy common sense. This limitation, stemming from the intrinsic data imbalance between text and video, is challenging to address due to the substantial cost of generating and annotating counterfactual data.
To address this, we introduce <strong>DualityForge</strong>, a novel counterfactual data synthesis framework that employs controllable, diffusion-based video editing to transform real-world videos into counterfactual scenarios.
By embedding structured contextual information into the video editing and QA generation processes, the framework automatically produces high‑quality QA pairs together with original–edited video pairs for contrastive training.
Based on this, we build <strong>DualityVidQA</strong>, a large-scale video dataset designed to reduce MLLM hallucinations.
Besides, to fully exploit the contrastive nature of our paired data, we propose <strong>D</strong>uality-<strong>N</strong>ormalized <strong>A</strong>dvantage <strong>Train</strong>ing (<strong>DNA-Train</strong>), a two-stage SFT-RL training regime where the RL phase incorporates <i>l</i><sub>1</sub> normalization of advantages for each real-counterfactual pair, thereby enabling a more stable and efficient policy optimization.
Experiments on <strong>DualityVidQA-Test</strong> demonstrate that our method substantially reduces model hallucinations on counterfactual videos, yielding a relative improvement of <strong>24.0%</strong> over the Qwen2.5-VL-7B baseline. Moreover, our approach achieves significant gains across both hallucination and general-purpose benchmarks, indicating strong generalization capability.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
<!-- Paper video. -->
<!-- <div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Video</h2>
<div class="publication-video">
<iframe src="https://www.youtube.com/embed/MrKrnHhk8IA?rel=0&showinfo=0"
frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
</div>
</div>
</div> -->
<!--/ Paper video. -->
</div>
</section>
<section class="">
<div class="container is-max-desktop">
<div>
<div class="is-centered has-text-centered">
<h2 class="title is-3">Framework Overview</h2>
</div>
<div class="hero-body" style="padding-bottom: 0; padding-left: 100px; padding-right: 100px">
<img src="./assets/images/framework.png" alt="framework" height="100%">
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<p>
<b>DualityVidQA Dataset.</b>We begin with a web-sourced real-video dataset and apply a framework integrating MLLMs, grounding and segmentation modules, and image/video editing models to synthesize counterfactual (CF) videos with targeted visual, semantic, and commonsense alterations. Each real-CF video pair is paired with MLLM-generated questions using carefully designed prompts. The dataset comprises three splits: DualityVidQA-SFT with real and counterfactual video-QA pairs (54K + 50K) for SFT; DualityVidQA-RL with 20K shared-question contrastive video-answer pairs (one question, two real/CF instances) for RL; and DualityVidQA-Test (600 pairs), which shares the same contrastive structure as DualityVidQA-RL and covers diverse counterfactual categories.
</p>
</div>
</div>
</div>
<!-- example video -->
<div class="is-centered has-text-centered">
<h3 class="title is-4">Example Video</h3>
</div>
<div class="hero-body has-text-centered">
<video class="video-center" width="60%" playsinline="" autoplay="autoplay" loop="loop" preload="" muted="">
<source src="./assets/videos/example_video.mp4" type="video/mp4">
</video>
</div>
<div class="is-centered has-text-centered">
<h3 class="title is-4">Edit Pipeline</h3>
</div>
<div style="padding-left: 100px; padding-right: 100px; padding-bottom: 0; padding-top: 1rem">
<img src="./assets/images/edit-pipeline.png" alt="edit_pipeline" height="100%">
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<p>
<!-- \caption{\textbf{Overview of DNA-Train framework.} We first perform SFT on our dual dataset to initialize the model. During RL, we sample a group of responses for both real and CF videos, compute their rewards based on task correctness, and calculate the $\ell_1$ norm of intra-group advantages. Finally, we normalize the advantages across the dual groups to ensure balanced gradients.}
-->
<b>Video Edit Pipeline </b> There are 3 pipeline are shown: 1. Visual Anomaly: we use opencv to edit the video in pixel-level. 2. Semantic Anomaly: we use VACE to edit the video in object-level. 3. Common Sense Anomaly: we use MLLM to generate the edit instruction, then use FLUX-Kontext to edit the first frame to end frame, finally use VACE to interpolate the video.
</p>
</div>
</div>
</div>
<div class="is-centered has-text-centered">
<h3 class="title is-4">DNA-Train Framework</h3>
</div>
<div style="padding-left: 100px; padding-right: 100px; padding-bottom: 0; padding-top: 1rem">
<img src="./assets/images/dna-train.png" alt="dna-train" height="100%">
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<p>
<!-- \caption{\textbf{Overview of DNA-Train framework.} We first perform SFT on our dual dataset to initialize the model. During RL, we sample a group of responses for both real and CF videos, compute their rewards based on task correctness, and calculate the $\ell_1$ norm of intra-group advantages. Finally, we normalize the advantages across the dual groups to ensure balanced gradients.}
-->
<b>DNA-Train framework.</b> We first perform SFT on our dual dataset to initialize the model. During RL, we sample a group of responses for both real and CF videos, compute their rewards based on task correctness, and calculate the l_1 norm of intra-group advantages. Finally, we normalize the advantages across the dual groups to ensure balanced gradients.
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="">
<div class="container is-max-desktop">
<div>
<div class="is-centered has-text-centered">
<h2 class="title is-3">Examples of LLM QA</h2>
</div>
<div class="hero-body">
<!-- <video width="100%" playsinline="" autoplay="autoplay" loop="loop" preload="" muted="">
<source src="./assets/images/example.png" type="video/mp4">
</video> -->
<img class="qa-image" src="./assets/images/example.png" alt="example">
</div>
</div>
</div>
</section>
<section class="">
<div class="container is-max-desktop">
<div class="is-centered has-text-centered">
<h2 class="title is-3">Dataset Statistics</h2>
</div>
<div class="columns is-multiline stats-grid">
<!-- Video Dataset Type -->
<div class="column is-12-tablet is-4-desktop">
<div class="box">
<h3 class="title is-6">Video Dataset Type Statistics</h3>
<div class="table-container">
<table class="table is-bordered is-narrow is-fullwidth is-size-7">
<thead>
<tr>
<th>Type</th><th>Count</th>
</tr>
</thead>
<tbody>
<tr><td>color</td><td>27353</td></tr>
<tr><td>replacement</td><td>9961</td></tr>
<tr><td>appearance</td><td>6092</td></tr>
<tr><td>disappear</td><td>5016</td></tr>
<tr><td>common sense</td><td>86746</td></tr>
<tr><th>All</th><th>133168</th></tr>
</tbody>
</table>
</div>
</div>
</div>
<!-- QA Type Frequency Overall -->
<div class="column is-12-tablet is-4-desktop">
<div class="box">
<h3 class="title is-6">QA Type Frequency Statistics (Overall)</h3>
<div class="table-container">
<table class="table is-bordered is-narrow is-fullwidth is-size-7">
<thead>
<tr>
<th>QA Type</th><th>Real Video</th><th>Counterfactual Video</th>
</tr>
</thead>
<tbody>
<tr><td>Multiple Choice</td><td>12210</td><td>10224</td></tr>
<tr><td>Open-Ended</td><td>42669</td><td>39776</td></tr>
<tr><th>All</th><th>54879</th><th>50000</th></tr>
</tbody>
</table>
</div>
</div>
</div>
<!-- Counterfactual Category -->
<div class="column is-12-tablet is-4-desktop">
<div class="box">
<h3 class="title is-6">Counterfactual Video Category Statistics</h3>
<div class="table-container">
<table class="table is-bordered is-narrow is-fullwidth is-size-7">
<thead>
<tr><th>Tag</th><th>Count</th></tr>
</thead>
<tbody>
<tr><td>causal reversal</td><td>158</td></tr>
<tr><td>counter physical</td><td>221</td></tr>
<tr><td>object/scene deformation</td><td>187</td></tr>
<tr><td>attribute change</td><td>33</td></tr>
<tr><th>All</th><th>599</td></tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div class="has-text-centered">
<h2 class="title">BibTeX</h2>
</div>
<div class="columns is-centered has-text-centered" style="padding-bottom: 1.5rem">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<pre><code>@article{huang2025taming,
title={Taming Hallucinations: Boosting MLLMs' Video Understanding via Counterfactual Video Generation},
author={Huang, Zhe and Wen, Hao and Hao, Aiming and Song, Bingze and Wu, Meiqi and Wu, Jiahong and Chu, Xiangxiang and Lu, Sheng and Wang, Haoqian},
journal={arXiv preprint arXiv:2512.24271},
year={2025}
}</code>
</pre>
</div>
</div>
</div>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<a class="icon-link"
href="">
<i class="fas fa-file-pdf"></i>
</a>
<a class="icon-link" href="https://github.com/Costwen/Ouroboros3D" class="external-link" disabled>
<i class="fab fa-github"></i>
</a>
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p style="text-align: center;">
The website template is borrowed from <a href="https://nerfies.github.io/" target="_blank">Nerfies</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>