-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
814 lines (714 loc) · 36.8 KB
/
index.html
File metadata and controls
814 lines (714 loc) · 36.8 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
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description" content="AdaptVision is an open-source model that leverages agentic visual tool use for dynamic visual token reduction">
<meta name="keywords" content="multimodal agent">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>AdaptVision</title>
<link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bulma@0.9.1/css/bulma.min.css">
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/4.5.2/css/bootstrap.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.1/css/all.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="https://cdn-icons-png.flaticon.com/512/954/954591.png">
<link href="https://fonts.googleapis.com/icon?family=Material+Icons" rel="stylesheet">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.1/js/all.min.js"></script>
<script type="module" src="https://gradio.s3-us-west-2.amazonaws.com/4.16.0/gradio.js"></script>
<script>
window.MathJax = {
tex: {
inlineMath: [['$', '$'], ['\\(', '\\)']],
displayMath: [['$$', '$$'], ['\\[', '\\]']]
},
svg: { fontCache: 'global' }
};
</script>
<!-- 从 CDN 引入 MathJax v3 -->
<script async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-svg.js"></script>
</head>
<style>
.expandable-card .card-text-container {
max-height: 200px;
overflow-y: hidden;
position: relative;
}
.expandable-card.expanded .card-text-container {
max-height: none;
}
.expand-btn {
position: relative;
display: none;
background-color: rgba(255, 255, 255, 0.8);
/* margin-top: -20px; */
/* justify-content: center; */
color: #510c75;
border-color: transparent;
}
.expand-btn:hover {
background-color: rgba(200, 200, 200, 0.8);
text-decoration: none;
border-color: transparent;
color: #510c75;
}
.expand-btn:focus {
outline: none;
text-decoration: none;
}
.expandable-card:not(.expanded) .card-text-container:after {
content: "";
position: absolute;
bottom: 0;
left: 0;
width: 100%;
height: 90px;
background: linear-gradient(rgba(255, 255, 255, 0.2), rgba(255, 255, 255, 1));
}
.expandable-card:not(.expanded) .expand-btn {
margin-top: -40px;
}
.card-body {
padding-bottom: 5px;
}
.vertical-flex-layout {
justify-content: center;
align-items: center;
height: 100%;
display: flex;
flex-direction: column;
gap: 5px;
}
.figure-img {
max-width: 100%;
height: auto;
}
.adjustable-font-size {
font-size: calc(0.5rem + 2vw);
}
.chat-history {
flex-grow: 1;
overflow-y: auto;
/* overflow-x: hidden; */
padding: 5px;
border-bottom: 1px solid #ccc;
margin-bottom: 10px;
}
#gradio pre {
background-color: transparent;
}
</style>
<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">AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://scholar.google.com/citations?user=Tlc4yaMAAAAJ&hl=zh-TW" style="color:#f68946;font-weight:normal;">Zichuan Lin<sup>*</sup></a>,
</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=8V7FwaUAAAAJ&hl=zh-CN" style="color:#008AD7;font-weight:normal;">Yicheng Liu<sup>*</sup></a>,
</span>
<span class="author-block">
<a href="" style="color:#F2A900;font-weight:normal;">Yang Yang</a>,
</span>
<span class="author-block">
<a href="" style="color:#f68946;font-weight:normal;">Lvfang Tao</a>,
</span>
<span class="author-block">
<a href="" style="color:#f68946;font-weight:normal;">Deheng Ye</a>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><b style="color:#5546f6; font-weight:normal">▶ </b> Tencent Hunyuan</b></span>
</div>
<div class="is-size-6 publication-authors">
<span class="author-block"><sup>*</sup>Equal Contribution</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/abs/2512.03794" 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>
<span class="link-block">
<a href="https://github.com/AdaptVision/AdaptVision" 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>
<!-- <span class="link-block">
<a href="https://huggingface.co/Mini-o3/datasets" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-database"></i>
</span>
<span>Dataset</span>
</a>
</span> -->
<span class="link-block">
<a href="https://huggingface.co/adaptvision/models" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-share-square"></i>
</span>
<span>Model</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<h4 class="subtitle has-text-centered">
AdaptVision is an open-source model that leverages agentic visual tool use for dynamic visual token reduction, achieving a sota-level accuracy-efficiency trade-off across multiple VQA benchmarks.
</h4>
<div style="text-align: center;">
<img id="teaser" width="80%" src="images/overall_performance.png">
</div>
</div>
</div>
</section>
<section class="section" style="background-color:#efeff081">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-six-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Vision-Language Models (VLMs) have achieved remarkable success in visual question answering tasks, but their reliance on large numbers of visual tokens introduces significant computational overhead. While existing efficient VLM approaches reduce visual tokens through fixed-ratio compression, they operate passively and lack the ability to adapt to varying task requirements. This motivates a fundamental question: Can VLMs autonomously determine the minimum number of visual tokens required for each sample? Inspired by human active vision mechanisms, we introduce AdaptVision, an efficient VLM paradigm that enables adaptive visual token acquisition through a coarse-to-fine approach. Our model initially processes compressed visual tokens from low-resolution images and selectively acquires additional visual information by invoking a bounding box tool to crop key regions when necessary. We train AdaptVision using a reinforcement learning framework that carefully balances accuracy and efficiency. Central to our approach is Decoupled Turn Policy Optimization (DTPO), which decouples the learning objective into two components: (1) tool learning, which optimizes correct tool utilization, and (2) accuracy improvement, which refines the generated responses to improve answer correctness. Based on this formulation, we further decouple advantage estimation by computing separate advantages for tokens associated with each objective. This formulation enables more effective optimization for AdaptVision compared to vanilla GRPO. Comprehensive experiments across multiple VQA benchmarks demonstrate that AdaptVision achieves superior performance while consuming substantially fewer visual tokens than state-of-the-art efficient VLM methods.
<ol type="1">
<li><b>Synergizing Visual Reasoning and Visual Token Compression</b>. <span style="font-size: 95%;">We introduce AdaptVision, a VLM framework that leverages visual tool use for dynamic token reduction.</span></li>
<li><b>Efficient Algorithm</b>. <span style="font-size: 95%;">We propose a Decoupled Turn Policy Optimization (DTPO) algorithm alongside a tailored reward function to enable the effective training of AdaptVision.</span></li>
<li><b>Performance</b>. <span style="font-size: 95%;">Extensive evaluation on multiple VQA benchmarks shows that AdaptVision achieves superior performance with substantially reduced visual token consumption compared to existing efficient VLM methods.</span></li>
<li><b>Open-source</b>. <span style="font-size: 95%;">All code, models, and training recipes are available to facilitate reproducibility and further research.</span></li>
<!-- <li><b>Challenging Dataset</b>. <span style="font-size: 95%;">We construct the <span style="color: #ff3860">Visual Probe Dataset</span>, a collection of thousands of challenging visual search problems designed for exploratory reasoning.</span></li> -->
<!-- <li><b>Diverse Multi-turn Trajectories for Cold-start</b>. <span style="font-size: 95%;">We develop an iterative data collection pipeline to obtain cold-start trajectories that exhibit <span style="color: #ff3860">diverse reasoning patterns, including depth-first search, trial-and-error, and goal maintenance</span>.</li>
<li><b>Test-time Turns Scaling</b>. <span style="font-size: 95%;">We propose an <b>over-turn masking</b> strategy that prevents penalization of responses exceeding the maximum turns during reinforcement learning, thereby balancing training-time efficiency with test-time scalability. Despite training with an upper bound of only six interaction turns, our model generates trajectories that naturally <span style="color: #ff3860">scale to tens of turns at inference time, with accuracy improving as the number of turns increases</span>.</li>
<li><b>Performance</b>. <span style="font-size: 95%;">Extensive experiments demonstrate that Mini-o3 produces rich reasoning patterns and deep thinking paths, effectively solving challenging visual problems, thereby achieving the <span style="color: #ff3860">state-of-the-art results on a variety of benchmarks</span> (e.g., VisualProbe, V* Bench, HR-Bench, MME-Realworld).</li>
<li><b>Open-source</b>. <span style="font-size: 95%;">All code, models, and datasets are available to facilitate reproducibility and further research.</li> -->
</ol>
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<!-- Results. -->
<div class="columns is-centered has-text-centered">
<div class="column is-six-fifths">
<h2 class="title is-3"><img id="painting_icon" width="3%" src="https://cdn-icons-png.flaticon.com/512/5379/5379860.png"> Learning Framework for AdaptVision</h2>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<div class="content has-text-justified">
<p>
<ul type="1">
<li>
<b>Framework Overview</b>
<div>
AdaptVision first processes a 1/4-resolution image. The model then decides whether to answer directly or invoke the bounding box tool to crop a high-resolution region for further analysis before generating the final answer.
</div>
<!-- <div>
AdaptVision is trained via reinforcement learning to maximize the outcome reward and tool reward.
The outcome reward $\mathcal{R}_{oc}$ provides a sequence-level feedback.
1) Accuracy reward uses an external LLM to judge the answer correctness.
2) Format reward enforces the instruction-following capability.
3) Balance reward prevents over reliance on tool calls.
\begin{equation}
\begin{split}
\mathcal{R}_{oc} = \mathcal{R}_{acc} + \mathcal{R}_{format} + \mathcal{R}_{balance}
\end{split}
\end{equation}
The tool reward $\mathcal{R}_{tool}$ measures the tool-use proficiency. It rewards correctly cropped regions and penalizes excessive cropping.
\begin{equation}
\begin{split}
\mathcal{R}_{tool} = \mathcal{R}_{crop} - \alpha \cdot \mathcal{R}_{area}
\end{split}
\end{equation}
</div> -->
<!-- <br> -->
<div style="text-align: center;">
<img id="demo" width="80%" src="images/framework.png">
</div>
<br>
</li>
<!-- <div>
We initially attempted to train the model with reinforcement learning alone, without cold-start supervised fine-tuning (SFT). However, the model tended to produce concise responses and trajectories with few turns. We attribute this behavior to the base model’s lack of exposure to long-horizon agentic trajectories during pretraining and instruction tuning (here, Qwen2.5-VL-7B-Instruct). To handle complex exploratory tasks, we thus employ cold-start SFT to activate multi-turn tool-use capabilities.
</div>
<br>
<div>
The cold-start data collection pipeline is shown in the above figure. To generate high-quality, diverse multi-turn trajectories, we prompt an existing VLM with in-context learning ability using a small set of manually crafted exemplars. The VLM is instructed to imitate the exemplars by iteratively producing a thought and an action at each turn. The loop terminates upon emitting a final answer or reaching a pre-defined turn limit. We retain only trajectories whose final answers are correct. Following this procedure, we collect approximately 6,000 cold-start trajectories from 6 exemplars.
</div>
<br> -->
<li>
<b>Reward Design</b>
<ul type="1">
<!-- <li> -->
1. Outcome Reward.
<span style="font-size: 100%;">
The outcome reward provides a sequence-level feedback.
1) Accuracy reward uses an external LLM to judge the answer correctness.
2) Format reward enforces the instruction-following capability.
3) Balance reward prevents over reliance on tool calls.
\begin{equation}
\begin{split}
\mathcal{R}_{oc} = \mathcal{R}_{acc} + \mathcal{R}_{format} + \mathcal{R}_{balance}
\end{split}
\end{equation}
</span>
<!-- </li> -->
<!-- <br> -->
<!-- <li> -->
2. Tool Reward.
<span style="font-size: 100%;">
The tool reward measures the tool-use proficiency. It rewards correctly cropped regions and penalizes excessive cropping.
\begin{equation}
\begin{split}
\mathcal{R}_{tool} = \mathcal{R}_{crop} - \alpha \cdot \mathcal{R}_{area}
\end{split}
\end{equation}
</span>
<!-- </li> -->
</ul>
</li>
<li>
<b>Decoupled Turn Policy Optimization (DTPO)</b>
<!-- <div>
AdaptVision first processes a 1/4-resolution image. The model then decides whether to answer directly or invoke the bounding box tool to crop a high-resolution region for further analysis before generating the final answer.
</div> -->
<ul type="1">
1. Balanced Optimization Objective:
<span style="font-size: 100%;">
DTPO decouples the policy loss by turns and normalize the contributions of tool and answer tokens separately.
This adjustment effectively resolves the under-optimization problem of tool tokens.
</span>
\begin{equation}
\mathcal{J}_{\text{GRPO}}(\theta)
= \mathbb{E}_{x, o_i}
\Bigg[
\frac{1}{G} \sum_{t=1}^{G} \frac{1}{N_i} \sum_{t=1}^{N_i} \mathcal{L}_{i,t}(\theta)
\Bigg]
= \mathbb{E}_{x, o_i}
\Bigg[
\underbrace{\frac{1}{G} \sum_{t=1}^{G} \frac{1}{N_i} \sum_{t=1}^{T_i} \mathcal{L}_{i,t}(\theta) }_{\textup{Tool Token}}
+ \underbrace{\frac{1}{G} \sum_{t=1}^{G} \frac{1}{N_i} \sum_{t=T_i+1}^{N_i} \mathcal{L}_{i,t}(\theta)}_{\textup{Answer Token}}
\Bigg].
\end{equation}
\begin{equation}
\mathcal{J}_{\text{DTPO}}(\theta) = \mathbb{E}_{x, o_i}
\Bigg[
\underbrace{\frac{1}{\sum_{i=1}^G T_i} \sum_{i=1}^G \sum_{t=1}^{T_i} \mathcal{L}_{i,t}(\theta)}_{\textup{Tool Token}}
+ \underbrace{\frac{1}{\sum_{i=1}^G (N_i - T_i)} \sum_{i=1}^G \sum_{t=T_i+1}^{N_i} \mathcal{L}_{i,t}(\theta) }_{\textup{Answer Token}}
\Bigg].
\end{equation}
<br>
2. Precise Credit Assignment:
<span style="font-size: 100%;">
DTPO decouples the advantage estimation by computing distinct advantages for tool and answer tokens,
rather than using a single advantage for the entire sequence.
</span>
\begin{equation}
A_{i,t}^{\text{GRPO}} = \frac{R_i - \text{mean}(\{R_i\}^G_{i=1})}{\text{std}(\{R_i\}^G_{i=1})} .
\end{equation}
\begin{gather}
A_{i,t}^{\text{DTPO}} =
\begin{cases}
A_{oc}^{(i)} + \lambda \cdot A_{tool}^{(i)}, & \textup{direct answer}, \\
A_{oc}^{(i)} + \lambda \cdot A_{tool}^{(i)} \cdot \mathbb{I}(1 \le t \le T_{i}) , & \textup{tool call},
\end{cases} \\
A_{tool}^{(i)} = \frac{\mathcal{R}_{tool}^{(i)} - \text{mean}(\{\mathcal{R}_{tool}^{(i)}\}^G_{i=1})}{\text{std}(\{\mathcal{R}_{tool}^{(i)}\}^G_{i=1})},
\quad \quad
A_{oc}^{(i)} = \frac{\mathcal{R}_{oc}^{(i)} - \text{mean}(\{\mathcal{R}_{oc}^{(i)}\}^G_{i=1})}{\text{std}(\{\mathcal{R}_{oc}^{(i)}\}^G_{i=1})} .
\end{gather}
</ul>
<br>
<div style="text-align: center;">
<img id="demo" width="80%" src="images/dtpo.png">
</div>
<br>
</li>
<!-- <ul type="1">
<li><b>Lower Down Max Pixels</b>. <span style="font-size: 100%;">The base model’s context length is constrained to 32K tokens. With the default image budget of roughly 12M pixels, the allowable number of interaction turns becomes severely limited by context, which hampers trial-and-error exploration on difficult tasks. To increase the feasible turn count per episode, we reduce the maximum pixels per image to 2M (or lower if necessary). This simple adjustment allows more turns to fit within the same context budget, improving solve rates on long-horizon problems.</span> </li>
<br>
<li><b>Over-turn Masking</b>. <span style="font-size: 100%;">In the vanilla GRPO setting, each question $q$ is passed to the policy model to generate a group of outputs $\{o_i\}_{i=1}^{G}$. Rewards $r$ are then computed based on the correctness of the responses. Notably, when a response hits the maximum number of turns or exceeds the context length limit, the reward is set to $0$, as no valid answer can be produced in such cases. Subsequently, we compute advantages $A$ by normalizing the rewards and update the policy using the GRPO optimization objective over mini-batches. In our implementation, we do not include KL or entropy regularization. Formally, the optimization objective is given by: -->
<!-- \begin{equation}
\begin{split}
\mathcal{J}_{GRPO}(\theta) = \mathbb{E}_{[q \sim \mathcal{D}, \{o_i\}_{i=1}^G \sim \pi_{\theta_{old}}(\cdot|q)]} \frac{1}{G}\sum_{i=1}^G \left(\min\left(\frac{\pi_\theta(o_i |q)}{\pi_{\theta_{old}}(o_i |q)} A_i, \text{clip} \left( \frac{\pi_\theta(o_i |q)}{\pi_{\theta_{old}}(o_i |q)}, 1 - \epsilon, 1 + \epsilon \right) A_i \right) \right)
\end{split}
\label{eq:GRPO}
\end{equation}
\begin{equation}
\begin{split}
A_i = \frac{r_i - mean(\{r_1,r_2,...,r_G\})}{std(\{r_1, r_2, ...,r_G\})}.
\end{split}
\label{eq:advantage}
\end{equation} -->
<!-- <div>
However, we observe that overlength responses --- those that hit the maximum number of turns or exceed the context length --- are assigned zero reward, which translates into negative advantages after normalization. In effect, such responses are penalized and discouraged throughout training.
</div>
<br>
<div style="text-align: center;">
<img id="demo" width="90%" src="images/turn_dist.png">
</div>
<br> -->
<!-- <div>
This design has two drawbacks. First, the correctness of overlength responses is inherently unknown; blunt penalization thus injects label noise into the return signal and can destabilize training. Second, for efficiency, the turn limit during training must remain modest (typically fewer than $10$ turns). As a consequence, overlength responses occur frequently --- exceeding $20\%$ at the beginning of training. In this regime, naïve penalization biases the model to answer prematurely, substantially suppressing the number of interaction turns (see above figure). This makes highly challenging tasks intractable and severely constrains the potential of test-time scaling.
</div>
<br>
<div style="text-align: center;">
<img id="demo" width="90%" src="images/overlength_mask.png">
</div>
<br> -->
<!-- <div>
To prevent the model from collapsing into an “answer earlier” strategy, we propose an overlength masking technique whose objective is to avoid penalizing overlength responses. The overall procedure is illustrated in the above figure. Concretely, in addition to the rewards $r$ and advantages $A$ defined as in vanilla GRPO, we introduce a completion mask $M$ that indicates whether a response terminates successfully. We then compute masked advantages $A_i'=M_i \cdot A_i$, so that overlength trajectories (with $M_i=0$) do not contribute negative learning signals. The modified objective, building on the vanilla GRPO, is summarized below, with the changes highlighted in red in the formula.
</div> -->
<!-- \begin{equation}
\begin{split}
\mathcal{J}^{\textcolor{red}{overlength}}_{GRPO}(\theta) & = \mathbb{E}_{[q \sim \mathcal{D}, \{o_i\}_{i=1}^G \sim \pi_{\theta_{old}}(\cdot|q)]} \\
\frac{1}{\textcolor{red}{\sum_i^{G}M_i}}\sum_{i=1}^G & \left(\min\left(\frac{\pi_\theta(o_i |q)}{\pi_{\theta_{old}}(o_i |q)} A_i \textcolor{red}{\cdot M_i}, \text{clip} \left( \frac{\pi_\theta(o_i |q)}{\pi_{\theta_{old}}(o_i |q)}, 1 - \epsilon, 1 + \epsilon \right) A_i\textcolor{red}{\cdot M_i} \right) \right)
\end{split}
\label{eq:new_GRPO}
\end{equation}
\begin{equation}
\begin{split}
\textcolor{red}{M_i = \mathcal{1}\{|o_i| <= C_{context}\} \cdot \mathcal{1}\{ \text{turn}(o_i) <= C_{turn}\}}.
\end{split}
\label{eq:new_advantage}
\end{equation} -->
<!-- <div>
Here, $|o_i|$ and $\text{turn}(o_i)$ denote the token length and the number of turns in response $o_i$, respectively. Moreover, because some responses are incomplete, we normalize the objective by the number of completed generations, $\sum_i^{G}M_i$, rather than by the total number of generations $G$.
</div>
<div>
With this technique, we mask out the loss for overlength responses, thereby removing any implicit penalty. Notably, although we adopt a relatively small upper bound on the number of turns during training, test-time trajectories can extend to dozens of rounds, with accuracy improving monotonically. The proposed overlength masking is thus essential for realizing the benefits of test-time scaling in the number of interaction turns, as illustrated in the above figure.
</div> -->
<!-- </span> -->
</ul>
</p>
</div>
</div>
</div>
</section>
<!-- <section class="section">
<div class="columns is-centered has-text-centered">
<div class="column is-six-fifths">
<h2 class="title is-3"><img id="painting_icon" width="3%" src="https://cdn-icons-png.flaticon.com/512/5886/5886212.png"> Visual Probe Dataset</h2>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<div class="content has-text-justified">
<p>
Hard instances are essential for encouraging reflective, trial-and-error reasoning during reinforcement learning. To this end, we construct a challenging visual search dataset, the <b>Visual Probe Dataset</b> (VisualProbe). It comprises 4,000 visual question–answer pairs for training and 500 pairs for testing, spanning three difficulty levels: easy, medium, and hard. Compared with prior visual search benchmarks, VisualProbe is characterized by:
<ul type="1">
<li><b>Small targets</b></li>
<li><b>Numerous distractor objects</b></li>
<li><b>High-resolution images</b></li>
</ul>
An example is illustrated in the below figure. These properties make the tasks substantially more demanding and naturally require iterative exploration and trial-and-error. Please check it out on
<a href="https://huggingface.co/Mini-o3/datasets">[HuggingFace Dataset]</a>.
<style>
table.GeneratedTable {
width: 100%;
background-color: #ffffff;
border-collapse: collapse;
border-width: 2px;
border-color: #c1c4c5;
border-style: solid;
color: #000000;
}
table.GeneratedTable td, table.GeneratedTable th {
border-width: 2px;
border-color: #9b9d9e;
border-style: solid;
padding: 3px;
}
table.GeneratedTable thead {
background-color: #6691ee;
}
</style>
</p>
</div>
</div>
</div>
<div style="text-align: center;">
<img id="demo" width="70%" src="images/visualprobe.png">
</div>
</section> -->
<section class="section">
<!-- Results. -->
<div class="columns is-centered has-text-centered">
<div class="column is-six-fifths">
<h2 class="title is-3">
<img id="painting_icon" width="3%" src="static/images/demo.jpg"> Demo
</h2>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<li>
<b>Tool-call Response</b>.
<span style="font-size: 100%;">
The down-sample model, while reducing visual token usage, fails to answer correctly due to insufficient information in the low-resolution image.
The vanilla model, using the original high-resolution image, yields a correct answer but at the cost of a large number of visual tokens.
In contrast, AdaptVision begins with the low-resolution image, analyzes the question and image, recognizes the informational inadequacy, and then intelligently invokes the tool to crop the most relevant region from the high-resolution image. By acquiring only this essential additional visual information, it produces an accurate answer while minimizing visual token consumption.
<br>
<br>
<centering>
<div style="text-align: center;">
<img id="demo" width="70%" src="images/toolcall_case_1.png">
<img id="demo" width="70%" src="images/toolcall_case_2.png">
</div>
</centering>
</span>
</li>
</div>
</div>
<br>
<br>
<div class="container is-max-desktop">
<div class="columns is-centered">
<li>
<b>Direct Answer</b>.
<span style="font-size: 100%;">
In scenarios where a low-resolution image provides enough information, AdaptVision correctly chooses to answer directly—matching the behavior of the Qwen2.5-VL Down-sample model.
<br>
<br>
<centering>
<div style="text-align: center;">
<img id="demo" width="70%" src="images/direct_case_1.png">
<img id="demo" width="70%" src="images/direct_case_2.png">
</div>
</centering>
</span>
</li>
</div>
</div>
</section>
<section class="section">
<div class="columns is-centered has-text-centered">
<div class="column is-six-fifths">
<h2 class="title is-3"><img id="painting_icon" width="3%" src="https://cdn-icons-png.flaticon.com/512/3515/3515174.png"> Performance</h2>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-4">AdaptVision achieves superior performance with substantially reduced visual token consumption compared to existing efficient VLM methods </h2>
<div>
<div class="columns is-centered">
<centering>
<div style="text-align: center;">
<img id="result" width="100%" src="images/main_result.png">
</div>
</centering>
</div>
<div class="table-notes">
<div>*Vanilla denotes the Qwen2.5-VL-7B-Instruct model.</div>
<div>*Down-Sample uses a 1/4-resolution image as input to the Vanilla model.</div>
</div>
<!-- <style>
table.perf-table {
border-collapse: collapse;
width: 95%;
text-align: center;
font-family: Arial, sans-serif;
}
table.perf-table caption {
caption-side: top;
font-weight: bold;
margin-bottom: 6px;
}
table.perf-table th, table.perf-table td {
border: 1px solid #ccc;
padding: 6px 8px;
}
table.perf-table thead th {
background: #f6f6f6;
}
/* 模拟 \rowcolor{Gray} */
.row-gray {
background: #eeeeee;
}
/* 左对齐第一列（Model 名称） */
.col-model {
text-align: left;
white-space: nowrap;
}
/* 分组列之间加粗边界，模拟 LaTeX 的竖线分组 */
.group-sep-left {
border-left: 2px solid #999 !important;
}
.group-sep-right {
border-right: 2px solid #999 !important;
}
/* 脚注样式 */
.table-notes {
font-size: 0.9em;
color: #444;
margin-top: 6px;
}
</style> -->
<!-- <table class="perf-table">
<caption>
Performance comparison with previous efficient VLM methods.
Vanilla denotes the Qwen2.5-VL-7B-Instruct model. Down-Sample uses a 1/4-resolution image as input to the Vanilla model.
``\#Token'' indicates the visual token consumption ratio relative to the vanilla model across all benchmarks.
``Avg.'' denotes the average performance relative to the vanilla model on all benchmarks.
</caption>
<thead>
<tr>
<th rowspan="2" class="group-sep-right">Model</th>
<th colspan="3" class="group-sep-right">VisualProbe</th>
<th rowspan="2" class="group-sep-right">V* Bench</th>
<th colspan="2" class="group-sep-right">HR-Bench</th>
<th rowspan="2">MME-Realworld</th>
</tr>
<tr>
<th>hard</th>
<th>medium</th>
<th class="group-sep-right">easy</th>
<th>4K</th>
<th class="group-sep-right">8K</th>
</tr>
</thead>
<tbody>
<tr>
<td class="col-model group-sep-right">GPT-4o</td>
<td>11.2</td>
<td>15.4</td>
<td class="group-sep-right">47.5</td>
<td class="group-sep-right">65.2</td>
<td>62.0</td>
<td class="group-sep-right">58.3</td>
<td>45.2</td>
</tr>
<tr>
<td class="col-model group-sep-right">LLaVA-OneVision</td>
<td>13.4</td>
<td>12.5</td>
<td class="group-sep-right">36.2</td>
<td class="group-sep-right">70.9</td>
<td>61.2</td>
<td class="group-sep-right">54.0</td>
<td>57.4</td>
</tr>
<tr>
<td class="col-model group-sep-right">Qwen2.5-VL-Instruct</td>
<td>23.9</td>
<td>26.0</td>
<td class="group-sep-right">39.1</td>
<td class="group-sep-right">75.5</td>
<td>68.2</td>
<td class="group-sep-right">62.7</td>
<td>57.3</td>
</tr>
<tr>
<td class="col-model group-sep-right">SEAL<sup>†</sup></td>
<td>–</td>
<td>–</td>
<td class="group-sep-right">–</td>
<td class="group-sep-right">75.4</td>
<td>–</td>
<td class="group-sep-right">–</td>
<td>–</td>
</tr>
<tr>
<td class="col-model group-sep-right">DyFo<sup>†</sup></td>
<td>–</td>
<td>–</td>
<td class="group-sep-right">–</td>
<td class="group-sep-right">81.2</td>
<td>–</td>
<td class="group-sep-right">–</td>
<td>–</td>
</tr>
<tr>
<td class="col-model group-sep-right">Chain-of-Focus<sup>†</sup></td>
<td>–</td>
<td>–</td>
<td class="group-sep-right">–</td>
<td class="group-sep-right">88.0</td>
<td>–</td>
<td class="group-sep-right">–</td>
<td>–</td>
</tr>
<tr>
<td class="col-model group-sep-right">Pixel Reasoner<sup>‡</sup></td>
<td>28.8</td>
<td>29.6</td>
<td class="group-sep-right">58.4</td>
<td class="group-sep-right">86.3</td>
<td>74.0</td>
<td class="group-sep-right">66.9</td>
<td>64.4</td>
</tr>
<tr>
<td class="col-model group-sep-right">DeepEyes<sup>‡</sup></td>
<td>35.1</td>
<td>29.8</td>
<td class="group-sep-right">60.1</td>
<td class="group-sep-right">83.3</td>
<td>73.2</td>
<td class="group-sep-right">69.5</td>
<td>64.0</td>
</tr>
<tr class="row-gray">
<td class="col-model group-sep-right">Mini-o3 (Ours)</td>
<td>48.0</td>
<td>50.4</td>
<td class="group-sep-right">67.0</td>
<td class="group-sep-right">88.2</td>
<td>77.5</td>
<td class="group-sep-right">73.3</td>
<td>65.5</td>
</tr>
</tbody>
</table> -->
<!-- <div class="table-notes">
<div>† The models only report the metric of Avg@1 and the model weights are not available.</div>
<div>‡ Re-evaluated using its official model and evaluation code to yield the metric of Avg@32.</div>
</div> -->
</div>
</div>
</div>
</section>
<section class="section" id="Acknowledgement">
<div class="container is-max-desktop content">
<h2 class="title">Acknowledgement</h2>
<p>
This website is adapted from <a
href="https://github.com/nerfies/nerfies.github.io">Nerfies</a>, <a
href="https://llava-vl.github.io/">LLaVA</a> and <a
href="https://mini-o3.github.io/">Mini-o3</a>, licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>. We thank the Qwen team for giving us access to their models, and open-source projects including VisionThink.
</p>
<p>
<b>Usage and License Notices</b>: The data, code and checkpoint is intended and licensed for research use only. They are also restricted to uses that follow the license agreement of Qwen and Gemini-2.5-Pro. The dataset is CC BY NC 4.0 (allowing only non-commercial use) and models trained using the dataset should not be used outside of research purposes.
</p>
</div>
</section>
</body>
</html>