Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
748 lines (730 loc) · 41.4 KB
/
Copy pathindex.html
File metadata and controls
748 lines (730 loc) · 41.4 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="description" content="CodeDance is a dynamic tool-integrated multimodal large language model for executable visual reasoning with adaptive tool use and reinforcement learning.">
<meta name="keywords" content="CodeDance, executable visual reasoning, multimodal large language models, MLLM, adaptive tool use, reinforcement learning, thinking with images">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="color-scheme" content="light">
<title>CodeDance: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning</title>
<link rel="canonical" href="https://codedance-vl.github.io/">
<link rel="icon" href="static/images/20251218-logo.png">
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/index.css?v=20261023-1">
<script src="static/js/site.js?v=20261023-1" defer></script>
<script src="static/js/mathjax.js?v=20261023-1" defer></script>
</head>
<body>
<a class="skip-link" href="#main-content">Skip to main content</a>
<main id="main-content" tabindex="-1">
<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">
<span class="has-text-danger has-text-weight-semibold">CodeDance</span>: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://openreview.net/profile?id=~Qi_Song9">Qi Song</a><sup>1,4♣</sup>,
</span>
<span class="author-block">
<a href="https://openreview.net/profile?id=%7EHonglin_Li2">Honglin Li</a><sup>2,4♣</sup>,
</span>
<span class="author-block">
<a href="https://yingchen001.github.io/">Yingchen Yu</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://www.zhouhaoyi.com/">Haoyi Zhou</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://openreview.net/profile?id=~Lin_Yang13">Lin Yang</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="https://songbai.site/">Song Bai</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://qi-she.net/">Qi She</a><sup>4</sup>
</span>
<br>
<span class="author-block">
<a href="https://speedinghzl.github.io/">Zilong Huang</a><sup>3♠</sup>
</span>
<span class="author-block">
<a href="https://yunqing-me.github.io/">Yunqing Zhao</a><sup>3♠</sup>
</span>
</div>
<br>
<div class="is-size-5 publication-authors">
<span class="author-block"><b>CVPR 2026 (Main Track), Denver</b></span>
</div>
<br>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Beihang University</span>
<span class="author-block"><sup>2</sup>Westlake University</span>
<span class="author-block"><sup>3</sup>ByteDance (SG)</span>
<span class="author-block"><sup>4</sup>ByteDance (CN)</span>
<br>
<span class="author-block"><sup>♣</sup>Equal Contribution</span>
<span class="author-block"><sup>♠</sup>Corresponding Authors</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2512.17312" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false"><path d="M6 3h8l4 4v14H6z M14 3v5h5 M9 12h6 M9 16h6"></path></svg></span>
<span>Paper</span>
</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/CodeDance-VL/CodeDance" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false"><path d="m8 7-5 5 5 5 M16 7l5 5-5 5 M14 4l-4 16"></path></svg></span>
<span>Code</span>
</a>
</span>
<!-- SFT Dataset Link. -->
<span class="link-block">
<a href="https://huggingface.co/datasets/Peterwatwec/CodeDance-SFT" target="_blank" class="external-link button is-normal is-rounded is-dark" rel="noopener noreferrer">
<span class="icon"><svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false"><ellipse cx="12" cy="5" rx="8" ry="3"></ellipse><path d="M4 5v14c0 4 16 4 16 0V5 M4 12c0 4 16 4 16 0"></path></svg></span>
<span>SFT Dataset</span>
</a>
</span>
<!-- RL Dataset Link. -->
<span class="link-block">
<a href="https://huggingface.co/datasets/Peterwatwec/CodeDance-RL" target="_blank" class="external-link button is-normal is-rounded is-dark" rel="noopener noreferrer">
<span class="icon"><svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false"><ellipse cx="12" cy="5" rx="8" ry="3"></ellipse><path d="M4 5v14c0 4 16 4 16 0V5 M4 12c0 4 16 4 16 0"></path></svg></span>
<span>RL Dataset</span>
</a>
</span>
<!-- Models Link. -->
<span class="link-block">
<a href="https://github.com/CodeDance-VL/CodeDance" target="_blank" class="external-link button is-normal is-rounded is-dark" rel="noopener noreferrer">
<span class="icon"><svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false"><path d="m12 2 9 5v10l-9 5-9-5V7z M3 7l9 5 9-5 M12 12v10"></path></svg></span>
<span>Models</span>
</a>
</span>
<!-- Demo Link. -->
<span class="link-block">
<a href="https://github.com/CodeDance-VL/CodeDance" target="_blank" class="external-link button is-normal is-rounded is-dark" rel="noopener noreferrer">
<span class="icon"><svg xmlns="http://www.w3.org/2000/svg" viewbox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false"><circle cx="12" cy="12" r="9"></circle><path d="m10 8 6 4-6 4z"></path></svg></span>
<span>Demo</span>
</a>
</span>
</div>
</div>
</div>
</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-2">Abstract</h2>
<div class="content has-text-justified">
<p>We introduce <span class="has-text-danger has-text-weight-semibold">CodeDance</span>, a <span class="has-text-info has-text-weight-semibold">dynamic tool-integrated multimodal large language model</span> that treats executable code as a general solver for visual reasoning. Instead of relying on rigid, text-only pipelines, CodeDance plans, composes, and executes visual–symbolic operations while rendering intermediate artifacts, enabling a transparent, <span class="has-text-info has-text-weight-semibold">self-checkable reasoning process</span>.</p>
<p><span class="has-text-danger has-text-weight-semibold">CodeDance</span> scales up multimodal tool-based reasoning by letting the model <span class="has-text-info has-text-weight-semibold">think, write code, execute it, and reflect</span> in a single loop. Rather than following fixed schemas, CodeDance dynamically decides when and how to invoke tools, orchestrates visual-symbolic operations (crop, draw, count, plot) in a sandbox, and uses intermediate visual evidence to guide subsequent reasoning over <span class="has-text-info has-text-weight-semibold">multiple interaction turns</span>.</p>
<p>Our design choice yields <span class="has-text-info has-text-weight-semibold">transparent, self-checkable solutions</span> to challenging visual search and reasoning tasks and offers a practical recipe for “thinking with images” via executable code and reinforcement learning.</p>
<p><span class="has-text-danger has-text-weight-semibold">Importantly</span>, we answer the question: "<b>Do we always need tools for visual reasoning?</b>" Our work demonstrates that <b>state-of-the-art MLLM reasoning performance does not necessarily require always-on tool use ability</b>.
In contrast, <span class="has-text-danger has-text-weight-semibold">the tool-integrated visual reasoning should be task-adaptive and difficulty-aware</span>. </p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-10">
<h2 class="title is-3 has-text-centered">Dynamic, <span class="has-text-danger has-text-weight-semibold">Difficulty-aware</span>, visual tool-calling</h2>
<div class="content has-text-justified">
<p><span class="has-text-danger has-text-weight-semibold">CodeDance</span> dynamically invokes tools <span class="has-text-info has-text-weight-semibold">according to task difficulty</span> to prevent both <span class="has-text-danger has-text-weight-semibold">tool underuse</span> (unable to invoke tools on challenging tasks) and <span class="has-text-danger has-text-weight-semibold">tool overuse</span> (redundant calls on easy tasks).
</p>
</div>
<figure class="research-figure"><a href="./static/images/teaser1_page-0001.jpg" class="figure-link" aria-label="View full-resolution figure: Comparison of tool underuse, tool overuse, and difficulty-adaptive tool invocation in CodeDance, with benchmark results."><img src="./static/images/teaser1_page-0001.jpg" alt="Comparison of tool underuse, tool overuse, and difficulty-adaptive tool invocation in CodeDance, with benchmark results." width="4459" height="2013" decoding="async">
</a><figcaption><b>Motivation and effectiveness of dynamic tool invocation in CodeDance.</b>
<b>Left:</b> qualitative examples show that both tool underuse (unable to invoke tools on challenging tasks) and tool overuse (redundant calls on easy tasks) lead to hallucinated reasoning, incorrect answers, and unnecessary complexity (more reasoning turns and longer rollout time), whereas CodeDance dynamically invokes tools according to task difficulty to obtain correct solutions.
<b>Right:</b> quantitative results show CodeDance-7B consistently surpasses all Qwen2.5-VL-7B baselines and <span class="has-text-danger has-text-weight-semibold">exceeds the 32B version and GPT-4o</span> on several tasks.</figcaption>
</figure>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-10">
<h2 class="title is-3 has-text-centered">Overview and Training Recipes</h2>
<div class="content has-text-justified">
<p><span class="has-text-danger has-text-weight-semibold">CodeDance</span> is a <span class="has-text-info has-text-weight-semibold"> dynamic tool-integrated MLLM</span> that leverages <span class="has-text-info has-text-weight-semibold">executable code</span> as a general solver for visual reasoning, enabling a transparent, <span class="has-text-info has-text-weight-semibold">self-checkable reasoning process</span>.
</p>
</div>
<a href="./static/images/method-1.png" class="figure-link" aria-label="View full-resolution figure: CodeDance pipeline: trajectory curation, supervised fine-tuning, reinforcement learning, and executable visual reasoning."><img src="./static/images/method-1.png" alt="CodeDance pipeline: trajectory curation, supervised fine-tuning, reinforcement learning, and executable visual reasoning." width="5550" height="4088" decoding="async" loading="lazy"></a><br>
</div>
</div>
<div class="columns is-centered">
<div class="column is-10">
<div class="content has-text-justified">
<p><b>Our training / inference pipeline consists of three stages:</b></p>
<ul>
<li>
<strong><span class="has-text-danger has-text-weight-semibold">Stage 1: Task-aware Cold-start via Supervised Fine-tuning.</span></strong> We construct a <span class="has-text-weight-semibold">34k high-quality</span> dataset of <em>executable</em> multi-turn trajectories to initialize the model before RL.
<ul>
<li>
<span class="has-text-info has-text-weight-semibold">Weak-to-strong filtering.</span> Aggregate public datasets (e.g., SA1B, GEOqa_plus, MMK12) and apply automatic filtering. We employ a weak model (Qwen2.5-VL-7B) to prune trivial cases, while a strong model further stratifies the remaining samples into medium and hard difficulty levels.</li>
<li>
<span class="has-text-info has-text-weight-semibold">Multi-turn atomic supervision.</span> We decompose hard cases into <span class="has-text-weight-semibold">verifiable executable trajectories</span> across three atomic categories:
<ul>
<li>
<span class="has-text-weight-semibold">Predefined visual operations.</span> Annotated via predefined workflows to ensure reproducibility and scalability. Visual outputs (e.g., cropped views) serve as evidence for stepwise reasoning.</li>
<li>
<span class="has-text-weight-semibold">Mathematical computation.</span> <span class="has-text-weight-semibold">Measurement, algebra, aggregation</span> executed via Python/NumPy/SymPy with <span class="has-text-weight-semibold">error filtering</span>. A <span class="has-text-weight-semibold">Reasoning Model</span> first produces chain-of-thought, which is segmented into a multi-step procedure; complex computational steps within the procedure are then translated into <span class="has-text-weight-semibold">executable code</span>.</li>
<li>
<span class="has-text-weight-semibold">Open-ended operations.</span> The model can execute <span class="has-text-weight-semibold">arbitrarily complex code</span> that helps solve the task (e.g., drawing, annotation), enabling flexible and task-adaptive tool use.</li>
</ul>
</li>
</ul>
</li>
<li>
<strong><span class="has-text-danger has-text-weight-semibold">Stage 2: Difficulty-adaptive Reinforcement Learning.</span></strong>
<div class="content has-text-justified">
<p>
We optimize with a <span class="has-text-info has-text-weight-semibold">composite reward</span> that integrates outcome and tool signals. Our <span class="has-text-info has-text-weight-semibold">two-level reward (Balanced Adaptive Tool-call)</span>, denoted as \( R_{\mathrm{BAT}} \), balances task difficulty with step-wise tool-call correctness:
</p>
<section class="math-scroll" tabindex="0" aria-label="Equation 1; scroll horizontally if needed">\[
R_{\mathrm{BAT}} = R_{\mathrm{seq}} + R_{\mathrm{turn}}.
\]</section>
<ul>
<li>
<p>
<span class="has-text-danger has-text-weight-semibold">Sequence-level.</span> Difficulty-aware incentives adapt to group accuracy to <span class="has-text-weight-semibold">discourage redundant calls</span> on easy problems and <span class="has-text-weight-semibold">promote exploration</span> on hard ones. Formally:
</p>
<p>
Here, \(N_{\mathrm{succ}}(\tau)\) and \(N_{\mathrm{total}}(\tau)\) denote successful and total tool calls in trajectory \(\tau\). The difficulty factor \(d\) shrinks when the group accuracy \(\mu_{\mathrm{acc}}\) is high (easy queries) and grows when \(\mu_{\mathrm{acc}}\) is low (hard queries), suppressing redundant tool invocations on easy cases while amplifying rewards for helpful tool use on challenging ones.
</p>
<section class="math-scroll" tabindex="0" aria-label="Equation 2; scroll horizontally if needed">\[
R_{\mathrm{seq}} =
\Big(0.5 + 0.5 \cdot \mathbb{I}_{R_{\mathrm{acc}}(\tau) > 0}\Big)
\cdot d \cdot \frac{N_{\mathrm{succ}}(\tau)}{N_{\mathrm{total}}(\tau)},
\quad
\]</section>
where
<section class="math-scroll" tabindex="0" aria-label="Equation 3; scroll horizontally if needed">\[
d = \sigma\!\big(\gamma (0.5 - \mu_{\mathrm{acc}})\big) - \delta,
\quad
\sigma(z) = \frac{1}{1 + e^{-z}}.
\]</section>
</li>
<li>
<p>
<span class="has-text-danger has-text-weight-semibold">Turn-level.</span> Immediate penalties for failed executions plus a batch-normalized advantage to provide dense correction. For each turn \(m\), failed code execution incurs an immediate penalty \(r_{\mathrm{turn},m} = -0.5\) (otherwise \(0\)). To capture long-term effects, we define the accumulated discounted return and advantage as:
</p>
<section class="math-scroll" tabindex="0" aria-label="Equation 4; scroll horizontally if needed">\[
G_{\mathrm{turn}}^{m} = r_{\mathrm{turn}}^{m} + \beta \cdot G_{\mathrm{turn}}^{m+1}, \qquad
A_{\mathrm{turn}}^{m} = \frac{G_{\mathrm{turn}}^{m} - \mu_{\mathrm{batch}}}{\sigma_{\mathrm{batch}}}.
\]</section>
<p>
This turn-level signal assigns credit to correct intermediate tool executions, discouraging reward hacking where the final answer is correct but tool use is erroneous.
</p>
<p>
The final advantage combines sequence- and turn-level signals:
</p>
<section class="math-scroll" tabindex="0" aria-label="Equation 5; scroll horizontally if needed">\[
A(\tau) = A_{\mathrm{seq}}\!\big(R_{\mathrm{acc}}, R_{\mathrm{format}}, R_{\mathrm{seq}}\big) + A_{\mathrm{turn}}\!\big(G_{\mathrm{turn}}\big).
\]</section>
</li>
</ul>
</div>
</li>
</ul>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-10">
<h2 class="title is-3 has-text-centered">Synthesis of <span class="has-text-danger has-text-weight-semibold">tool-integrated visual</span> trajectories</h2>
<div class="content has-text-justified">
<p>To initialize CodeDance before RL, we curate a high-quality cold-start dataset of executable multimodal trajectories. Following the pipeline described in our methodology and appendix, we apply weak-to-strong filtering to stratify difficulty and then build multi-turn atomic supervision. The first figure shows trajectory synthesis; the second shows an example curated trajectory. Select a figure to view it at full resolution.</p>
</div>
<div class="trajectory-grid">
<figure class="research-figure"><a href="./static/images/data-syn.png" class="figure-link" aria-label="View full-resolution figure: Pipeline for synthesizing executable reasoning trajectories from visual questions."><img src="./static/images/data-syn.png" alt="Pipeline for synthesizing executable reasoning trajectories from visual questions." width="2266" height="2502" decoding="async" loading="lazy">
</a><figcaption>Trajectory synthesis pipeline.</figcaption></figure><figure class="research-figure"><a href="./static/images/data-image.png" class="figure-link" aria-label="View full-resolution figure: Example of a curated trajectory with code execution and intermediate visual evidence."><img src="./static/images/data-image.png" alt="Example of a curated trajectory with code execution and intermediate visual evidence." width="2146" height="2024" decoding="async" loading="lazy">
</a><figcaption>Example of a curated executable reasoning trajectory.</figcaption></figure>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-10">
<h2 class="title is-3 has-text-centered"> Experiments on <span class="has-text-danger has-text-weight-semibold">General & Reasoning</span> Benchmarks</h2>
<div class="content has-text-justified">
<p>We conduct systematic evaluations across visual search, math, counting, and chart understanding benchmarks. CodeDance consistently outperforms text-only and schema-driven baselines and, in several settings, surpasses larger open-source and advanced closed-source models.</p>
<p class="scroll-hint" aria-hidden="true">Swipe or scroll horizontally to see all columns.</p>
<section class="table-container" tabindex="0" aria-label="Results on visual counting, visual search, and general reasoning benchmarks. Scroll horizontally to view all columns.">
<table class="table is-striped is-hoverable is-fullwidth">
<caption>Experiment results on benchmarks including Counting, Visual Search, and General datasets.</caption>
<thead>
<tr>
<th rowspan="2" scope="col" class="benchmark-model">Model</th>
<th colspan="2" class="has-text-centered" scope="colgroup">Visual Counting</th>
<th colspan="3" class="has-text-centered" scope="colgroup">Visual Search</th>
<th colspan="2" class="has-text-centered" scope="colgroup">General</th>
</tr>
<tr>
<th scope="col">CountBenchQA</th>
<th scope="col">PixmoCount</th>
<th scope="col">V*Bench</th>
<th scope="col">HR4K</th>
<th scope="col">HR8K</th>
<th scope="col">ChartQA</th>
<th scope="col">CharXiv</th>
</tr>
</thead>
<tbody>
<!-- Closed-Source MLLMs -->
<tr>
<td colspan="8" class="has-text-centered has-text-weight-bold">Closed-Source MLLMs</td>
</tr>
<tr>
<th scope="row">GPT-4o</th>
<td>87.9</td>
<td>-</td>
<td>67.5</td>
<td>65.0</td>
<td>59.6</td>
<td>86.7</td>
<td>47.1</td>
</tr>
<!-- Open-Source MLLMs -->
<tr>
<td colspan="8" class="has-text-centered has-text-weight-bold">Open-Source MLLMs</td>
</tr>
<tr>
<th scope="row">LLaVA-OneVision-7B</th>
<td>82.3</td>
<td>54.4</td>
<td>72.7</td>
<td>68.5</td>
<td>60.0</td>
<td>80.4</td>
<td>27.1</td>
</tr>
<tr>
<th scope="row">LLaVA-OneVision-72B</th>
<td>-</td>
<td>60.7</td>
<td>73.8</td>
<td>66.3</td>
<td>60.9</td>
<td>83.7</td>
<td>-</td>
</tr>
<tr>
<th scope="row">InternVL2.5-8B</th>
<td>55.9</td>
<td>-</td>
<td>73.7</td>
<td>72.0</td>
<td>65.5</td>
<td>82.8</td>
<td>37.2</td>
</tr>
<tr>
<th scope="row">InternVL3-8B</th>
<td>80.3</td>
<td>-</td>
<td>70.2</td>
<td>70.5</td>
<td>70.0</td>
<td>86.1</td>
<td>38.3</td>
</tr>
<tr>
<th scope="row">InternVL3-78B</th>
<td>-</td>
<td>-</td>
<td>76.4</td>
<td>75.5</td>
<td>67.3</td>
<td>89.7</td>
<td>46.0</td>
</tr>
<tr>
<th scope="row">Qwen2.5-VL-72B</th>
<td>93.6</td>
<td>62.3</td>
<td>84.8</td>
<td>79.4</td>
<td>76.3</td>
<td>89.5</td>
<td>49.7</td>
</tr>
<tr>
<th scope="row">Qwen2.5-VL-32B</th>
<td>87.8</td>
<td>56.0</td>
<td>85.9</td>
<td>74.8</td>
<td>71.6</td>
<td>-</td>
<td>47.6</td>
</tr>
<tr>
<th scope="row">Qwen2.5-VL-7B</th>
<td>76.5</td>
<td>50.4</td>
<td>76.4</td>
<td>69.0</td>
<td>66.0</td>
<td>86.3</td>
<td>42.1</td>
</tr>
<!-- Open-Source MLLMs with Tools -->
<tr>
<td colspan="8" class="has-text-centered has-text-weight-bold">Open-Source MLLMs with Tools</td>
</tr>
<tr>
<th scope="row">Pixel Reasoner-7B</th>
<td>-</td>
<td>-</td>
<td>84.3</td>
<td>72.9</td>
<td>66.9</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<th scope="row">DeepEyes-7B<sup>†</sup>
</th>
<td>80.4</td>
<td>57.2</td>
<td>90.4</td>
<td>74.8</td>
<td>71.9</td>
<td>78.2</td>
<td>-</td>
</tr>
<tr>
<th scope="row">Thyme-VL-7B</th>
<td>84.8<sup>†</sup>
</td>
<td>-</td>
<td>82.2</td>
<td>77.0</td>
<td>72.0</td>
<td>86.1</td>
<td>44.2<sup>†</sup>
</td>
</tr>
<!-- Ours -->
<tr class="has-background-light results-row">
<th scope="row">CodeDance-7B</th>
<td>91.2</td>
<td>77.1</td>
<td>84.8</td>
<td>75.2</td>
<td>72.3</td>
<td>87.5</td>
<td>44.1</td>
</tr>
<tr class="has-background-light">
<th scope="row"><i>Δ vs. Qwen2.5-VL-7B</i></th>
<td class="has-text-success has-text-weight-bold">↑19.2%</td>
<td class="has-text-success has-text-weight-bold">↑53.0%</td>
<td class="has-text-success has-text-weight-bold">↑11.0%</td>
<td class="has-text-success has-text-weight-bold">↑9.0%</td>
<td class="has-text-success has-text-weight-bold">↑9.5%</td>
<td class="has-text-success has-text-weight-bold">↑1.4%</td>
<td class="has-text-success has-text-weight-bold">↑4.7%</td>
</tr>
</tbody>
</table>
</section>
<br>
<br>
<p class="scroll-hint" aria-hidden="true">Swipe or scroll horizontally to see all columns.</p>
<section class="table-container" tabindex="0" aria-label="Results on mathematical reasoning benchmarks. Scroll horizontally to view all columns.">
<table class="table is-striped is-hoverable is-fullwidth">
<caption>Experiment results on math benchmarks.</caption>
<thead>
<tr>
<th scope="col">Model</th>
<th scope="col">MathVision</th>
<th scope="col">MathVista</th>
<th scope="col">MathVerse</th>
<th scope="col">WeMath</th>
</tr>
</thead>
<tbody>
<tr>
<th scope="row">GPT-4o</th>
<td>36.5</td>
<td>63.4</td>
<td>35.3</td>
<td>44.2</td>
</tr>
<tr>
<th scope="row">Qwen2.5-VL-72B</th>
<td>38.1</td>
<td>74.8</td>
<td>57.6</td>
<td>-</td>
</tr>
<tr>
<th scope="row">R1-OneVision-7B<sup>†</sup>
</th>
<td>29.9</td>
<td>64.1</td>
<td>40.0</td>
<td>-</td>
</tr>
<tr>
<th scope="row">R1-VL-7B<sup>†</sup>
</th>
<td>24.7</td>
<td>63.5</td>
<td>40.0</td>
<td>-</td>
</tr>
<tr>
<th scope="row">InternVL2.5-8B</th>
<td>22.0</td>
<td>64.4</td>
<td>39.5</td>
<td>23.9</td>
</tr>
<tr>
<th scope="row">LLaVA-OV-7B</th>
<td>18.4</td>
<td>63.2</td>
<td>26.2</td>
<td>17.3</td>
</tr>
<tr>
<th scope="row">Qwen2.5-VL-7B</th>
<td>25.0</td>
<td>68.1</td>
<td>45.1</td>
<td>35.4</td>
</tr>
<tr>
<th scope="row">DeepEyes-7B</th>
<td>26.6</td>
<td>70.1</td>
<td>47.3</td>
<td>38.9</td>
</tr>
<tr class="has-background-light">
<th scope="row">CodeDance-7B (Ours)</th>
<td>29.6</td>
<td>70.3</td>
<td>46.8</td>
<td>39.6</td>
</tr>
</tbody>
</table>
<br>
</section>
<p class="scroll-hint" aria-hidden="true">Swipe or scroll horizontally to see all columns.</p>
<section class="table-container" tabindex="0" aria-label="Reward-design ablation: accuracy and average reasoning turns. Scroll horizontally to view all columns.">
<table class="table is-striped is-hoverable is-fullwidth">
<caption>Experiment results on reward design.</caption>
<thead>
<tr>
<th scope="col">Components</th>
<th scope="col">CountBench<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">PixmoCount<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">MathVision<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">MathVerse<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">V*<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">HR4K<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">HR8K<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
<th scope="col">Avg.<br><span class="is-size-7 has-text-weight-normal">Acc. / Turns</span>
</th>
</tr>
</thead>
<tbody>
<tr>
<th scope="row">SFT Cold-Start (w/o RL)</th>
<td>85.3<br><span class="is-size-7">1.2749</span>
</td>
<td>66.9<br><span class="is-size-7">1.3902</span>
</td>
<td>23.0<br><span class="is-size-7">2.8388</span>
</td>
<td>41.4<br><span class="is-size-7">2.1904</span>
</td>
<td>82.7<br><span class="is-size-7">2.0052</span>
</td>
<td>72.1<br><span class="is-size-7">1.1713</span>
</td>
<td>67.1<br><span class="is-size-7">1.0875</span>
</td>
<td>62.6<br><span class="is-size-7">1.7083</span>
</td>
</tr>
<tr>
<th scope="row">RL with R<sub>acc</sub> + R<sub>format</sub>
</th>
<td>88.4<br><span class="is-size-7">1.0200</span>
</td>
<td>71.2<br><span class="is-size-7">1.0170</span>
</td>
<td>26.0<br><span class="is-size-7">2.1086</span>
</td>
<td>
<b>46.5</b><br><span class="is-size-7 has-text-success">1.9569</span>
</td>
<td>82.7<br><span class="is-size-7 has-text-success">1.1728</span>
</td>
<td>73.4<br><span class="is-size-7 has-text-success">1.0413</span>
</td>
<td>69.0<br><span class="is-size-7 has-text-success">1.0375</span>
</td>
<td>65.3<br><span class="is-size-7 has-text-success">1.3363</span>
</td>
</tr>
<tr>
<th scope="row">+ R<sub>DeepEyes</sub>
</th>
<td>85.1<br><span class="is-size-7 has-text-danger">2.5960</span>
</td>
<td>64.4<br><span class="is-size-7 has-text-danger">2.5341</span>
</td>
<td>25.2<br><span class="is-size-7 has-text-danger">3.2270</span>
</td>
<td>44.0<br><span class="is-size-7 has-text-danger">2.5190</span>
</td>
<td>
<b>83.3</b><br><span class="is-size-7 has-text-danger">2.0000</span>
</td>
<td>
<b>74.6</b><br><span class="is-size-7 has-text-danger">2.0888</span>
</td>
<td>68.4<br><span class="is-size-7 has-text-danger">2.0525</span>
</td>
<td>63.6<br><span class="is-size-7 has-text-danger">2.4311</span>
</td>
</tr>
<tr class="has-background-light">
<th scope="row">+ R<sub>BAT</sub> (Ours)</th>
<td>
<b>89.0</b><br><span class="is-size-7 has-text-success">1.0000</span>
</td>
<td>
<b>72.5</b><br><span class="is-size-7 has-text-success">1.0000</span>
</td>
<td>
<b>27.0</b><br><span class="is-size-7 has-text-success">2.0461</span>
</td>
<td>46.3<br><span class="is-size-7">2.1662</span>
</td>
<td>82.7<br><span class="is-size-7">1.2094</span>
</td>
<td>73.8<br><span class="is-size-7">1.2251</span>
</td>
<td>
<b>69.4</b><br><span class="is-size-7">1.1950</span>
</td>
<td>
<b>65.8</b><br><span class="is-size-7">1.4060</span>
</td>
</tr>
</tbody>
</table>
</section>
<p class="is-size-7 has-text-grey">Note: † denotes results reported from official papers.</p>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-10">
<h2 class="title is-3 has-text-centered">
<span class="has-text-danger has-text-weight-semibold">Key Findings: </span>Emergent Behaviors during RL</h2>
<div class="content has-text-justified">
<p>
Throughout the RL process, we consistently observe <span class="has-text-info has-text-weight-semibold">novel and surprising reasoning trajectories</span> (below) that <span class="has-text-danger has-text-weight-semibold">go beyond the atomic supervision</span> provided during SFT.
These findings point toward the <span class="has-text-danger has-text-weight-semibold">scalability</span> of code as a general reasoning medium, and we empirically study the potential in our work.
</p>
</div>
<a href="./static/images/emergent_behavior.jpg" class="figure-link" aria-label="View full-resolution figure: Examples of emergent tool transfer and novel tool composition during reinforcement learning."><img src="./static/images/emergent_behavior.jpg" alt="Examples of emergent tool transfer and novel tool composition during reinforcement learning." width="4330" height="3968" decoding="async" loading="lazy">
</a>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-10">
<h2 class="title is-3 has-text-centered">Training <span class="has-text-danger has-text-weight-semibold">scalability</span> and Test-time Scaling (TTS)</h2>
<div class="content has-text-justified">
<p>
We show that without task-specific fine-tuning, CodeDance exhibits <span class="has-text-info has-text-weight-semibold">emergent capabilities</span> beyond supervised primitives: <span class="has-text-weight-semibold">novel tool invocations</span>, <span class="has-text-weight-semibold">unseen tool compositions</span>, and <span class="has-text-weight-semibold">cross-task transfer</span>, demonstrating strong generalization at inference.
<span class="has-text-danger has-text-weight-semibold">We additionally validate the potential of scalability</span> across four main dimensions: <b>(1)</b> amount of training data, <b>(2)</b> scale of the foundation model, <b>(3)</b> max test-time reasoning turns and <b>(4)</b> number of RL steps.
</p>
</div>
<a href="./static/images/scaling.png" class="figure-link" aria-label="View full-resolution figure: CodeDance scaling experiments varying training data, model size, inference turns, and reinforcement learning steps."><img src="./static/images/scaling.png" alt="CodeDance scaling experiments varying training data, model size, inference turns, and reinforcement learning steps." width="2203" height="2302" decoding="async" loading="lazy">
</a>
</div>
</div>
</div>
</section>
<!-- BibTeX citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
If you find our work useful and inspiring to your research, please consider citing our work:
<pre tabindex="0" aria-label="BibTeX citation; scroll horizontally to read the full entry"><code>@article{Song2025codedance,
title={CodeDance: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning},
author={Song, Qi and Li, Honglin and Yu, Yingchen and Zhou, Haoyi and Yang, Lin and Bai, Song and She, Qi and Huang, Zilong and Zhao, Yunqing},
journal={arXiv preprint arXiv:2512.17312},
year={2025}
}</code></pre>
</div>
</section>
</main>
<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="https://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
<p>
This project page is constructed using the wonderful template provided by <a href="https://nerfies.github.io/">Nerfies</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>