-
Notifications
You must be signed in to change notification settings - Fork 19
Expand file tree
/
Copy pathindex.html
More file actions
705 lines (622 loc) · 33.6 KB
/
index.html
File metadata and controls
705 lines (622 loc) · 33.6 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<!-- Primary Meta Tags -->
<!-- TODO: Replace with your paper title and author names -->
<meta name="title" content="Native Parallel Reasoner">
<!-- TODO: Write a compelling 150-160 character description of your research -->
<meta name="description" content="We introduce the Native Parallel Reasoner (NPR), a scalable framework for constructing models that intrinsically reason in parallelism. NPR learns adaptive decomposition and aggregation policies through a teacher-free pipeline combining self-distilled parallel Supervised Fine-Tuning (SFT) with Native Parallel Reinforcement Learning (RL). This approach allows the model to optimize its own branching strategies directly from experience within a shared computation graph, preserving its native reasoning style while maximizing exploration efficiency. Across eight diverse reasoning benchmarks, NPR achieves decisive gains: self-distilled data outperform prior teacher-generated corpora by 10.1%, and our Parallel RL stage improves over direct RL baselines by 3.0%. Crucially, NPR delivers up to 4.6× inference acceleration over autoregressive baselines and exhibits genuine, non-simulated parallel reasoning behaviors. We release the complete NPR ecosystem, including code, models, and robust inference infrastructure, to democratize access to next-generation parallel cognitive architectures.">
<!-- TODO: Add 5-10 relevant keywords for your research area -->
<meta name="keywords" content="Parallel Reasoning, Reinforcement Learning, Self-Distillation, Large Language Models, AI Research">
<!-- TODO: List all authors -->
<meta name="author" content="Tong Wu, Yang Liu, Jun Bai, Zixia Jia, Shuyi Zhang, Ziyong Lin, Yanting Wang, Song-Chun Zhu, Zilong Zheng">
<meta name="robots" content="index, follow">
<meta name="language" content="English">
<!-- Twitter -->
<meta name="twitter:card" content="summary_large_image">
<!-- TODO: Replace with your lab/institution Twitter handle -->
<meta name="twitter:site" content="@YOUR_TWITTER_HANDLE">
<!-- TODO: Replace with first author's Twitter handle -->
<meta name="twitter:creator" content="@AUTHOR_TWITTER_HANDLE">
<!-- TODO: Same as paper title above -->
<meta name="twitter:title" content="PAPER_TITLE">
<!-- TODO: Same as description above -->
<meta name="twitter:description" content="BRIEF_DESCRIPTION_OF_YOUR_RESEARCH_CONTRIBUTION_AND_FINDINGS">
<!-- TODO: Same as social preview image above -->
<meta name="twitter:image" content="https://YOUR_DOMAIN.com/assets/images/social_preview.png">
<meta name="twitter:image:alt" content="PAPER_TITLE - Research Preview">
<!-- Academic/Research Specific -->
<meta name="citation_title" content="PAPER_TITLE">
<meta name="citation_author" content="FIRST_AUTHOR_LAST, FIRST_AUTHOR_FIRST">
<meta name="citation_author" content="SECOND_AUTHOR_LAST, SECOND_AUTHOR_FIRST">
<meta name="citation_publication_date" content="2024">
<meta name="citation_conference_title" content="CONFERENCE_NAME">
<meta name="citation_pdf_url" content="https://YOUR_DOMAIN.com/assets/pdfs/paper.pdf">
<!-- Additional SEO -->
<meta name="theme-color" content="#2563eb">
<meta name="msapplication-TileColor" content="#2563eb">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="default">
<!-- Preconnect for performance -->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link rel="preconnect" href="https://ajax.googleapis.com">
<link rel="preconnect" href="https://documentcloud.adobe.com">
<link rel="preconnect" href="https://cdn.jsdelivr.net">
<!-- TODO: Replace with your paper title and authors -->
<title>Native Parallel Reasoner</title>
<!-- Favicon and App Icons -->
<link rel="icon" type="image/x-icon" href="assets/images/nlco_icon.png">
<link rel="apple-touch-icon" href="assets/images/nlco_icon.png">
<!-- Critical CSS - Load synchronously -->
<link rel="stylesheet" href="assets/css/bulma.min.css">
<link rel="stylesheet" href="assets/css/index.css">
<!-- Non-critical CSS - Load asynchronously -->
<link rel="preload" href="assets/css/bulma-carousel.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="assets/css/bulma-slider.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="assets/css/fontawesome.all.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<!-- Fallback for browsers that don't support preload -->
<noscript>
<link rel="stylesheet" href="assets/css/bulma-carousel.min.css">
<link rel="stylesheet" href="assets/css/bulma-slider.min.css">
<link rel="stylesheet" href="assets/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
</noscript>
<!-- Fonts - Optimized loading -->
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap" rel="stylesheet">
<!-- Defer non-critical JavaScript -->
<script defer src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="assets/js/fontawesome.all.min.js"></script>
<script defer src="assets/js/bulma-carousel.min.js"></script>
<script defer src="assets/js/bulma-slider.min.js"></script>
<script defer src="assets/js/index.js"></script>
<!-- Structured Data for Academic Papers -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "ScholarlyArticle",
"headline": "PAPER_TITLE",
"description": "BRIEF_DESCRIPTION_OF_YOUR_RESEARCH_CONTRIBUTION_AND_FINDINGS",
"author": [
{
"@type": "Person",
"name": "FIRST_AUTHOR_NAME",
"affiliation": {
"@type": "Organization",
"name": "INSTITUTION_NAME"
}
},
{
"@type": "Person",
"name": "SECOND_AUTHOR_NAME",
"affiliation": {
"@type": "Organization",
"name": "INSTITUTION_NAME"
}
}
],
"datePublished": "2024-01-01",
"publisher": {
"@type": "Organization",
"name": "CONFERENCE_OR_JOURNAL_NAME"
},
"url": "https://YOUR_DOMAIN.com/YOUR_PROJECT_PAGE",
"image": "https://YOUR_DOMAIN.com/assets/images/social_preview.png",
"keywords": ["KEYWORD1", "KEYWORD2", "KEYWORD3", "machine learning", "computer vision"],
"abstract": "FULL_ABSTRACT_TEXT_HERE",
"citation": "BIBTEX_CITATION_HERE",
"isAccessibleForFree": true,
"license": "https://creativecommons.org/licenses/by/4.0/",
"mainEntity": {
"@type": "WebPage",
"@id": "https://YOUR_DOMAIN.com/YOUR_PROJECT_PAGE"
},
"about": [
{
"@type": "Thing",
"name": "RESEARCH_AREA_1"
},
{
"@type": "Thing",
"name": "RESEARCH_AREA_2"
}
]
}
</script>
<!-- Website/Organization Structured Data -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "NLCo Lab, Beijing Institute for General Artificial Intelligence (BIGAI)",
"url": "https://github.com/bigai-nlco",
"logo": "https://github.com/bigai-nlco/Native-Parallel-Reasoner/tree/main/assets/images/nlco_icon.png",
}
</script>
</head>
<body>
<!-- Scroll to Top Button -->
<button class="scroll-to-top" onclick="scrollToTop()" title="Scroll to top" aria-label="Scroll to top">
<i class="fas fa-chevron-up"></i>
</button>
<!-- More Works Dropdown -->
<div class="more-works-container">
<button class="more-works-btn" onclick="toggleMoreWorks()" title="View More Works from Our Lab">
<i class="fas fa-flask"></i>
More Works
<i class="fas fa-chevron-down dropdown-arrow"></i>
</button>
<div class="more-works-dropdown" id="moreWorksDropdown">
<div class="dropdown-header">
<h4>More Works from Our Lab</h4>
<button class="close-btn" onclick="toggleMoreWorks()">
<i class="fas fa-times"></i>
</button>
</div>
<div class="works-list">
<a href="https://arxiv.org/abs/2502.18890" class="work-item" target="_blank">
<div class="work-info">
<h5>TokenSwift: Lossless Acceleration of Ultra Long Sequence Generation</h5>
<!-- <p>Brief description of the work and its main contribution.</p> -->
<span class="work-venue">ICML 2025</span>
</div>
<i class="fas fa-external-link-alt"></i>
</a>
<a href="https://arxiv.org/abs/2506.08672" class="work-item" target="_blank">
<div class="work-info">
<h5>RuleReasoner: Reinforced Rule-based Reasoning via Domain-aware Dynamic Sampling</h5>
<!-- <p>Brief description of the work and its main contribution.</p> -->
<span class="work-venue">Preprint</span>
</div>
<i class="fas fa-external-link-alt"></i>
</a>
<a href="https://arxiv.org/abs/2508.19594" class="work-item" target="_blank">
<div class="work-info">
<h5>Understanding and Leveraging the Expert Specialization of Context Faithfulness in Mixture-of-Experts LLMs</h5>
<!-- <p>Brief description of the work and its main contribution.</p> -->
<span class="work-venue">EMNLP 2025</span>
</div>
<i class="fas fa-external-link-alt"></i>
</a>
<a href="https://arxiv.org/abs/2506.11603" class="work-item" target="_blank">
<div class="work-info">
<h5>TongSearch-QR: Reinforced Query Reasoning for Retrieval</h5>
<!-- <p>Brief description of the work and its main contribution.</p> -->
<span class="work-venue">EMNLP 2025</span>
</div>
<i class="fas fa-external-link-alt"></i>
</a>
<a href="https://aclanthology.org/2025.acl-long.1113/" class="work-item" target="_blank">
<div class="work-info">
<h5>Look Both Ways and No Sink: Converting LLMs into Text Encoders without Training</h5>
<!-- <p>Brief description of the work and its main contribution.</p> -->
<span class="work-venue">ACL 2025</span>
</div>
<i class="fas fa-external-link-alt"></i>
</a>
</div>
</div>
</div>
<main id="main-content">
<section class="hero">
<div class="hero-body">
<div class="container is-fluid">
<!-- <div class="container is-max-desktop"> -->
<div class="columns is-centered">
<div class="column has-text-centered">
<!-- TODO: Replace with your paper title -->
<h1 class="title is-1 publication-title">Native Parallel Reasoner: Reasoning in Parallelism<br> via Self-Distilled Reinforcement Learning</h1>
<div class="is-size-5 publication-authors">
<!-- TODO: Replace with your paper authors and their personal links -->
<span class="author-block">
<a href="https://wutong4012.github.io" target="_blank">Tong Wu</a><sup>*†</sup>,</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/yangliu1998" target="_blank">Yang Liu</a><sup>*</sup>,</span>
<span class="author-block">
<a href="https://ba1jun.github.io/" target="_blank">Jun Bai</a><sup>*</sup>,
</span>
<span class="author-block">
<a href="https://jzxxx.github.io" target="_blank">Zixia Jia</a><sup>†</sup>,
</span>
<span class="author-block">
<a href="FIFTH AUTHOR PERSONAL LINK" target="_blank">Shuyi Zhang</a>,
</span>
<span class="author-block">
<a href="https://aclanthology.org/people/ziyong-lin" target="_blank">Ziyong Lin</a>,
</span>
<span class="author-block">
<a href="https://github.com/NoaneWang" target="_blank">Yanting Wang</a>,
</span>
<br>
<span class="author-block">
<a href="https://zhusongchun.net" target="_blank">Song-Chun Zhu</a>,
</span>
<span class="author-block">
<a href="https://zilongzheng.github.io" target="_blank">Zilong Zheng</a><sup>†✉</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<!-- TODO: Replace with your institution and conference/journal info -->
<span class="author-block">NLCo Lab, Beijing Institute for General Artificial Intelligence (BIGAI)</span>
<!-- TODO: Remove this line if no equal contribution -->
<span class="eql-cntrb"><small><br>*Core Contributors</small></span>
<span class="eql-cntrb"><small> †Project Leaders</small></span>
<span class="eql-cntrb"><small> ✉Correspondence to: zlzheng(at)bigai.ai</small></span>
</div>
<div class="column has-text-centered">
<!-- TODO: Update with your arXiv paper ID -->
<span class="link-block">
<a href="https://arxiv.org/abs/2512.07461" 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>
<!-- TODO: Replace with your GitHub repository URL -->
<span class="link-block">
<a href="https://github.com/bigai-nlco/Native-Parallel-Reasoner" 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>
<!-- Model -->
<span class="link-block">
<a href="https://huggingface.co/bigai-NPR" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon" style="vertical-align: middle; font-size: 20px;">🤗</span>
<span style="vertical-align: middle;">Model</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Teaser video-->
<!-- <section class="hero teaser">
<div class="container is-max-desktop">
<h2 class="title is-2 has-text-centered">
🎬 Demo
</h2>
<hr>
<div class="hero-body">
<video poster="" id="tree" autoplay controls muted loop height="100%" preload="metadata">
<source src="assets/videos/case1.mp4" type="video/mp4">
</video>
<h2 class="subtitle has-text-centered">
NPR demonstrates its parallel reasoning process on a complex problem, showcasing its ability to decompose tasks and aggregate solutions efficiently.
</h2>
</div>
</div>
</section> -->
<section class="hero is-small">
<div class="hero-body">
<div class="container">
<h2 class="title is-2 has-text-centered", style="margin-bottom: 0.25rem;">
🎬 Demo
</h2>
<div id="results-carousel" class="carousel results-carousel">
<div class="item item-video1">
<!-- TODO: Add poster image for better preview -->
<video poster="" id="video1" controls muted loop height="100%" preload="metadata">
<!-- Your video file here -->
<source src="assets/videos/case1.mp4" type="video/mp4">
</video>
</div>
<div class="item item-video2">
<video poster="" id="video2" controls muted loop height="100%" preload="metadata">
<source src="assets/videos/case2.mp4" type="video/mp4">
</video>
</div>
</div>
<h2 class="subtitle has-text-centered">
NPR demonstrates its parallel reasoning process on a complex problem, showcasing its ability to<br> decompose tasks and aggregate solutions efficiently.
</h2>
</div>
</div>
</section>
<!-- End teaser video -->
<script>
document.addEventListener("DOMContentLoaded", function () {
// Initialize carousel (disable autoScroll!)
var carousels = bulmaCarousel.attach('#results-carousel', {
autoplay: false,
pauseOnHover: false,
navigation: true
});
// Get carousel instance
var carousel = carousels[0];
// Grab all videos
const videos = document.querySelectorAll("#results-carousel video");
// For each video, bind "ended" event
videos.forEach((video, index) => {
video.addEventListener("ended", () => {
// Move to next slide when the current video ends
carousel.next();
// Find next video's DOM element
const nextIndex = (index + 1) % videos.length;
const nextVideo = videos[nextIndex];
// Wait a moment (carousel transition), then play next
setTimeout(() => {
nextVideo.play();
}, 300); // depends on your carousel transition duration
});
});
// Auto-play the first video when page loads
videos[0].play();
});
</script>
<section class="section is-medium">
<div class="container">
<h2 class="title is-2 has-text-centered">
🔎 Why Should Reasoning Go Wider?
</h2>
<hr>
<div class="columns is-centered">
<div class="column is-10">
<div class="content is-medium" style="text-align: justify;">
<p>
The rapid development of super-scale Large Language Models (LLMs) like Gemini 3 and GPT-5 has fundamentally changed the focus of AI research.
We are moving beyond simply achieving <strong>semantic fluency</strong> toward enabling <strong>deep, multi-step agentic reasoning</strong>.
</p>
<p>
While current methods allow for <strong>"deeper"</strong> test-time scaling to solve complex problems sequentially, the new dominant requirement
for advanced agentic AI is <strong>"wider"</strong> reasoning capacity, i.e., <strong>parallel reasoning</strong>, enabling the model to explore diverse trajectories in parallel
rather than being limited to a single, linear thought process.
</p>
</div>
<div class="column is-centered has-text-centered">
<iframe width="100%" height="500px" src="https://www.youtube.com/embed/oxzs2KExNiE?si=ixzdrNf1fp-ySw0C" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>
</div>
</div>
</div>
</div>
</section>
<section class="section is-medium hero is-light">
<div class="container">
<h2 class="title is-2 has-text-centered">
📌 TL;DR
</h2>
<hr>
<div class="columns is-centered">
<div class="column is-10">
<div class="content is-medium" style="text-align: justify;">
<p>
We introduce <strong>Native Parallel Reasoner (NPR)</strong>, a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities.
NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: 1) a <strong>self-distilled</strong> progressive training paradigm that
transitions from "cold-start" format discovery to strict topological constraints without external supervision; 2) a novel <strong>Parallel-Aware Policy Optimization (PAPO)</strong>
algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and 3) a robust <strong>NPR Engine</strong>
that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains
of up to 24.5% and inference speedups up to 4.6x. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100% genuine parallel execution, establishing a new standard
for self-evolving, efficient, and scalable agentic reasoning.
</p>
</div>
<figure class="image is-centered">
<img src="assets/images/teaser.png" alt="NPR Overview" style="max-width: 100%; height: auto; display: block; margin: 0 auto;">
<figcaption class="has-text-centered is-size-6 mt-2">
Figure 1：Native Parallel Reasoner (NPR) transforms a base model from sequential chain-of-thought (CoT) to native parallel reasoning via a self-distilled progressive training paradigm.
Compared with previous SoTA, NPR achieves high reasoning accuracy, genuine parallelism and token acceleration. The illustrated results are evaluated on the AIME25 benchmark.
</figcaption>
</figure>
</div>
</div>
</div>
</section>
<!-- Paper abstract -->
<!-- <section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
We introduce <strong>Native Parallel Reasoner (NPR)</strong>, a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities.
NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: 1) a <strong>self-distilled</strong> progressive training paradigm that
transitions from ``cold-start'' format discovery to strict topological constraints without external supervision; 2) a novel <strong>Parallel-Aware Policy Optimization (PAPO)</strong>
algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and 3) a robust <strong>NPR Engine</strong>
that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains
of up to 24.5% and inference speedups up to 4.6x. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100% genuine parallel execution, establishing a new standard
for self-evolving, efficient, and scalable agentic reasoning.
</p>
</div>
</div>
</div>
</div>
</section> -->
<!-- End paper abstract -->
<section class="section hero">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">🧩 Method Overview</h2>
<div class="content has-text-justified">
<!-- TODO: Replace with your paper abstract -->
<p>
NPR adpots a three-stage curriculum, as illustrated in <strong>Figure 2</strong>, to progressively induce, ground, and amplify this capability:
(1) <strong>NPR-ZERO</strong>, which utilizes reinforcement learning to induce a parallel formatting schema without external supervision;
(2) <strong>NPR-BETA</strong>, which stabilizes the model's parallel primitives through supervised fine-tuning on self-distilled data; and
(3) <strong>NPR</strong>, which employs a novel parallel-aware RL (PAPO) algorithm to robustly optimize the native parallel reasoning performance.
</p>
</div>
<figure class="image is-centered">
<img src="assets/images/npr-overview.png" alt="NPR Overview" style="max-width: 100%; height: auto; display: block; margin: 0 auto;">
<figcaption class="has-text-centered is-size-6 mt-2">
Figure 2：An overview of the NPR training framework.
</figcaption>
</figure>
</div>
</div>
</div>
</section>
<section class="section hero">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">✨ Parallel-Aware Policy Optimization (PAPO)</h2>
<div class="content has-text-justified">
<p>
Parallel-Aware Policy Optimization (PAPO) (<strong>Figure 3</strong>) effectively leverages RL to enhance parallel reasoning, leading to more generalized agentic behavior.
It builds on the SFT foundation with four key modifications to stabilize training and enforce parallel semantics:
</p>
<ol>
<li>
<strong>NPR-Engine Rollouts:</strong> We utilize our custom NPR-Engine to strictly enforce the Map–Process–Reduce flow, ensuring all generated trajectories have valid parallel structure.
</li>
<li>
<strong>Structural Filtering:</strong> We apply schema-level filtering during rollout to guarantee high-quality, structurally correct samples enter optimization.
</li>
<li>
<strong>Batch-level Normalization:</strong> To counteract the variance collapse caused by filtering, we adopt batch-level advantage normalization, stabilizing the training process.
</li>
<li>
<strong>On-Policy Objective:</strong> To preserve crucial gradients on special tokens (which control parallel semantics) and avoid the instability of PPO's importance-sampling, we eliminate reweighting and adopt a strict on-policy objective, which both stabilizes and speeds up training.
</li>
</ol>
</div>
<figure class="image is-centered">
<img src="assets/images/PAPO.png" alt="PAPO" style="max-width: 100%; height: auto; display: block; margin: 0 auto;">
<figcaption class="has-text-centered is-size-6 mt-2">
Figure 3：Comparison of GRPO-style RL and Parallel-Aware Policy Optimization (PAPO).
</figcaption>
</figure>
</div>
</div>
</div>
</section>
<section class="section hero">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3 has-text-centered">🔥 Key Results: NPR Performance Advantages</h2>
<div class="content has-text-justified">
<p>
NPR significantly surpasses strong baselines across reasoning benchmarks (See <strong>Table 1</strong>), stem from three major architectural advantages:
</p>
<ul>
<li>
<strong>1. Training Data Advantage:</strong> Replacing sequential data with our NPR self-distillation corpus consistently improves solution correctness. The overall score increases by +8.9 points (50.1 to 59.0), with maximum gains observed on ZebraLogic (+15.9).
</li>
<li>
<strong>2. Parallel SFT Advantage:</strong> Switching to Parallel SFT eliminates the restrictive inter-step dependency of sequential methods, enabling flexible task decomposition. This results in a net performance gain and increased overall robustness (58.2 to 59.0).
</li>
<li>
<strong>3. Parallel RL Advantage:</strong> Applying PAPO effectively amplifies high-reward solution modes through tailored parallel optimization. This leads to a substantial overall boost in average performance from 62.0 to 65.0, with AIME24 improving by +6.2.
</li>
</ul>
</div>
<figure class="image is-centered">
<img src="assets/images/npr-results.png" alt="板块描述图片" style="max-width: 100%; height: auto; display: block; margin: 0 auto;">
<figcaption class="has-text-centered is-size-6 mt-2">
Table 1：Performance of sequential and parallel reasoners on reasoning benchmarks.
</figcaption>
</figure>
</div>
</div>
</div>
</section>
<section class="section hero">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3 has-text-centered">🛡️ Stable Parallel Reasoning Triggering</h2>
<div class="content has-text-justified">
<p> We compare the parallel reasoning trigger rate between our NPR model and the Multiverse baseline (See <strong>Table 2</strong>). This rate measures a model's propensity to generate simultaneous, non-sequential thought structures during problem-solving. </p>
<ul>
<li>
<strong>Multiverse (Inconsistent):</strong> The Multiverse model, which relies on a progressive sequential-to-parallel training paradigm, demonstrates poor and <strong>highly inconsistent</strong> performance.
</li>
<li>
<strong>NPR (Holistic & Consistent):</strong> In sharp contrast, our NPR model, trained with an end-to-end parallel pipeline (SFT and RL), achieves a <strong>perfect 100.0%</strong> parallel reasoning trigger rate across all evaluated datasets.
</li>
</ul>
</div>
<figure class="image is-centered">
<img src="assets/images/npr-parallel_rate.png" alt="板块描述图片" style="max-width: 100%; height: auto; display: block; margin: 0 auto;">
<figcaption class="has-text-centered is-size-6 mt-2">
Table 2：Comparison of parallel reasoning trigger rates between NPR and MultiVerse.
</figcaption>
</figure>
</div>
</div>
</div>
</section>
<section class="section hero">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3 has-text-centered">🚀 Extreme Inference Acceleration</h2>
<div class="content has-text-justified">
<p> We evaluate token throughput and acceleration rate against Multiverse and autoregressive (AR) baselines. As shown in <strong>Table 3</strong>, NPR consistently achieves the strongest efficiency and substantial speedups: </p>
<ul>
<li>
<strong>Superior Overall Efficiency:</strong> NPR consistently outperforms <strong>Multiverse</strong> by <strong>1.3× to 2.4×</strong> and surpasses AR baselines across all five evaluation benchmarks.
</li>
<li>
<strong>Scaling with Difficulty:</strong> NPR's efficiency advantage increases with task difficulty, validating our hypothesis that NPR excels when simultaneous exploration of multiple solution paths is required.
</li>
<li>
<strong>Maximum Speedup:</strong> The highest acceleration rates are observed on the most challenging tasks: <strong>AIME25 (4.6×)</strong> and <strong>HMMT25 (4.1×)</strong>, significantly higher than simpler tasks like AMC23 (2.9×).
</li>
</ul>
</div>
<figure class="image is-centered">
<img src="assets/images/npr-efficiency.png" alt="板块描述图片" style="max-width: 100%; height: auto; display: block; margin: 0 auto;">
<figcaption class="has-text-centered is-size-6 mt-2">
Table 3：Evaluation results of tokens per second (TPS) and speedup ratio. The speedup ratio (denoted as Speedup) is calculated by comparing with sequential reasoning baseline.
</figcaption>
</figure>
</div>
</div>
</div>
</section>
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div class="bibtex-header">
<h2 class="title">How to cite</h2>
<button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX to clipboard">
<i class="fas fa-copy"></i>
<span class="copy-text">Copy</span>
</button>
</div>
<pre id="bibtex-code"><code>@misc{nativeparallelreasonerreasoning,
title={Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning},
author={Tong Wu and Yang Liu and Jun Bai and Zixia Jia and Shuyi Zhang and Ziyong Lin and Yanting Wang and Song-Chun Zhu and Zilong Zheng},
year={2025},
eprint={2512.07461},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2512.07461},
}</code></pre>
</div>
</section>
<!--End BibTex citation -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
© 2025 <a href="https://github.com/bigai-nlco" target="_blank">BIGAI NLCo Lab</a>. All rights reserved. Website template credit to <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Statcounter tracking code -->
<!-- You can add a tracker to track page visits by creating an account at statcounter.com -->
<!-- End of Statcounter Code -->
</body>
</html>