Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
530 lines (484 loc) · 29.3 KB
/
Copy pathindex.html
File metadata and controls
530 lines (484 loc) · 29.3 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="description"
content="AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent.">
<meta name="keywords" content="AgentArk, Multi-Agent Systems, LLM Distillation, Process Reward Model, Reasoning, GRPO, Multi-Agent Debate">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="theme-color" content="#6d4692">
<title>AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent</title>
<meta property="og:title" content="AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent">
<meta property="og:description" content="A framework that distills multi-agent debate dynamics into a single LLM, lifting single-agent reasoning by +4.8% on average across 120 experiments.">
<meta property="og:type" content="website">
<link rel="stylesheet" href="./media/static/css/index.css">
<link rel="preload" href="./media/agentark-poster.webp" as="image" fetchpriority="high">
<link rel="icon" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'%3E%3Ctext y='.9em' font-size='90'%3E%F0%9F%9B%B8%3C/text%3E%3C/svg%3E">
<script defer src="./media/static/js/index.js"></script>
</head>
<body id="top">
<a class="skip-link" href="#main">Skip to content</a>
<!-- ============================ STICKY NAV ============================ -->
<nav class="topnav" aria-label="Page sections">
<div class="nav-inner">
<a href="#top" class="brand">
<span class="brand-emoji" aria-hidden="true">🛸</span>
<span class="brand-name">AgentArk</span>
</a>
<details class="mobile-nav">
<summary>Contents <svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="m6 9 6 6 6-6"/></svg></summary>
<div class="mobile-links">
<a href="#video">Video</a>
<a href="#abstract">Overview</a>
<a href="#method">Method</a>
<a href="#results">Results</a>
<a href="#findings">Findings</a>
<a href="#analysis">Analysis</a>
<a href="#bibtex">BibTeX</a>
</div>
</details>
<div class="links">
<a href="#video">Video</a>
<a href="#abstract">Overview</a>
<a href="#method">Method</a>
<a href="#results">Results</a>
<a href="#findings">Findings</a>
<a href="#analysis">Analysis</a>
<a href="#bibtex">BibTeX</a>
</div>
</div>
</nav>
<!-- ============================ HERO ============================ -->
<main id="main">
<header class="hero-section">
<div class="hero-inner">
<div class="venue-lockup">
<img class="venue-logo" src="./media/neurips.webp" width="640" height="288"
alt="Neural Information Processing Systems" fetchpriority="high">
<div class="venue-label">
<span class="venue-kicker">Accepted at</span>
<span class="venue-name">NeurIPS 2026</span>
</div>
</div>
<div class="title-row">
<h1 class="paper-title">
<span class="paper-name">AgentArk</span>
<span class="paper-subtitle">Distilling Multi-Agent Intelligence<br class="title-break"> into a Single LLM Agent</span>
</h1>
</div>
<p class="hero-lead">Multi-agent intelligence. Single-agent efficiency.</p>
<div class="author-list">
<span class="author-chip"><a href="https://www.linkedin.com/in/yinyi-luo-5b0805324">Yinyi Luo</a><sup>1</sup></span>
<span class="author-chip"><a href="https://ahren09.github.io/">Yiqiao Jin</a><sup>3</sup></span>
<span class="author-chip"><a href="https://weichen-yu.github.io/">Weichen Yu</a><sup>1</sup></span>
<span class="author-chip"><a href="https://scholar.google.com/citations?user=h7HjebkAAAAJ">Mengqi Zhang</a><sup>2</sup></span>
<span class="author-chip"><a href="https://faculty.cc.gatech.edu/~srijan/">Srijan Kumar</a><sup>3</sup></span>
<span class="author-chip"><a href="https://xxlya.github.io/">Xiaoxiao Li</a><sup>5</sup></span>
<span class="author-chip"><a href="https://www.linkedin.com/in/weijie-xu-936b23101/">Weijie Xu</a><sup>4</sup><span class="corr">*</span></span>
<span class="author-chip">Xin Chen<sup>4</sup></span>
<span class="author-chip"><a href="https://jd92.wang/">Jindong Wang</a><sup>2</sup><span class="corr">*</span></span>
</div>
<div class="affiliations">
<span><sup>1</sup>Carnegie Mellon University</span>
<span><sup>2</sup>William & Mary</span>
<span><sup>3</sup>Georgia Institute of Technology</span>
<span><sup>4</sup>Amazon</span>
<span><sup>5</sup>University of British Columbia</span>
</div>
<div class="corr-note"><span class="corr">*</span> Corresponding authors</div>
<div class="cta-row">
<a href="https://arxiv.org/abs/2602.03955" class="cta cta-paper" target="_blank" rel="noopener">
<svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="M14 3H6a1 1 0 0 0-1 1v16a1 1 0 0 0 1 1h12a1 1 0 0 0 1-1V8zM14 3v5h5M8 12h8M8 16h6"/></svg><span>Read the paper</span>
</a>
<a href="https://github.com/AIFrontierLab/AgentArk" class="cta cta-code" target="_blank" rel="noopener">
<svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="m8 7-5 5 5 5m8-10 5 5-5 5m-3-13-2 20"/></svg><span>Code</span>
</a>
<a href="./media/AgentArk_Poster.pdf" class="cta cta-poster" target="_blank" rel="noopener">
<svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><rect x="3" y="3" width="18" height="18" rx="2"/><path d="m3 16 5-5 4 4 3-3 6 6M8 7h.01"/></svg><span>Poster</span>
</a>
<a href="#video" class="cta cta-video">
<svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="m8 4 13 8-13 8z"/></svg><span>Watch video</span>
</a>
<a href="#bibtex" class="cta cta-outline">
<svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="M10 5H3v8h7V5Zm0 8c0 4-3 6-6 6M21 5h-7v8h7V5Zm0 8c0 4-3 6-6 6"/></svg><span>BibTeX</span>
</a>
</div>
</div>
</header>
<!-- ============================ VIDEO ============================ -->
<section class="video-section" id="video" aria-labelledby="video-title">
<div class="video-heading">
<h2 id="video-title">The idea in one minute</h2>
<span class="video-duration">01:00 <span aria-hidden="true">/</span> Research overview</span>
</div>
<video class="research-video" controls playsinline preload="none"
poster="./media/agentark-poster.webp" width="1920" height="1080"
aria-label="AgentArk: one-minute research overview"
aria-describedby="video-description">
<source src="./media/agentark.mp4" type="video/mp4">
Your browser does not support embedded video.
<a href="./media/agentark.mp4">Open the AgentArk video.</a>
</video>
<div class="video-footer">
<p id="video-description">From multi-agent debate to a single, more capable reasoning agent.</p>
<a href="./media/agentark.mp4" target="_blank" rel="noopener">Open video <span aria-hidden="true">↗</span></a>
</div>
</section>
<!-- ============================ HEADLINE METRICS ============================ -->
<div class="metric-strip">
<div class="metric-strip-grid">
<div class="metric-cell">
<span class="num">+4.8%</span>
<span class="label">Avg. accuracy lift over single-agent baseline</span>
</div>
<div class="metric-cell tone-teal">
<span class="num">120</span>
<span class="label">Experiments across Qwen3, Gemma 3, Llama 3</span>
</div>
<div class="metric-cell tone-warm">
<span class="num">3</span>
<span class="label">Distillation strategies: R-SFT · DA · PAD</span>
</div>
<div class="metric-cell tone-rose">
<span class="num">~342K</span>
<span class="label">Distillation questions, ~2M reasoning trajectories</span>
</div>
</div>
</div>
<!-- ============================ ABSTRACT ============================ -->
<section class="section" id="abstract">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">01 · Overview</div>
<h2 class="section-title">Abstract</h2>
</div>
<div class="abstract-card">
<p>
While large language model (LLM) multi-agent systems achieve superior reasoning performance through iterative debate, practical deployment is limited by their high computational cost and error propagation. This paper proposes <strong>AgentArk</strong>, a novel framework to distill multi-agent dynamics into the weights of a <em>single</em> model, effectively transforming explicit test-time interactions into implicit model capabilities. This equips a single agent with the intelligence of multi-agent systems while remaining computationally efficient. Specifically, we investigate three hierarchical distillation strategies across various models, tasks, scaling, and scenarios: reasoning-enhanced fine-tuning; trajectory-based augmentation; and process-aware distillation. By shifting the burden of computation from inference to training, the distilled models preserve the efficiency of one agent while exhibiting strong reasoning and self-correction performance of multiple agents. They further demonstrate enhanced robustness and generalization across diverse reasoning tasks. We hope this work can shed light on future research on efficient and robust multi-agent development. Our code is at
<a href="https://github.com/AIFrontierLab/AgentArk">github.com/AIFrontierLab/AgentArk</a>.
</p>
</div>
</div>
</section>
<!-- ============================ TEASER ============================ -->
<section class="section alt-bg">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">02 · Motivation</div>
<h2 class="section-title">From a debating crowd to a single thinker</h2>
<p class="section-lead">
Multi-agent debate produces strong reasoning but is slow and brittle. AgentArk compresses the dynamics of an agent ensemble into the weights of one model — preserving the collective behavior at single-agent inference cost.
</p>
</div>
<div class="fig-full">
<a class="figure-link" href="./media/figures/teaser.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: AgentArk teaser: distilling multi-agent reasoning into a single agent">
<img src="./media/figures/teaser.png" alt="AgentArk teaser: distilling multi-agent reasoning into a single agent" width="1391" height="624" loading="lazy" decoding="async">
</a>
<p class="caption">
AgentArk distills the reasoning capability of a multi-agent system into a single agent, so that one unit imitates the collective thinking process with boosted performance and a fraction of the inference cost.
</p>
</div>
</div>
</section>
<!-- ============================ KEY CONTRIBUTIONS ============================ -->
<section class="section">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">03 · Contributions</div>
<h2 class="section-title">What AgentArk delivers</h2>
</div>
<div class="grid-3">
<div class="feature-card">
<span class="feature-icon-wrap" aria-hidden="true">01</span>
<h4>First MAS-distillation study</h4>
<p>To our knowledge, the first comprehensive framework that explores and compares multiple strategies for distilling multi-agent reasoning into a single model.</p>
</div>
<div class="feature-card tone-teal">
<span class="feature-icon-wrap" aria-hidden="true">02</span>
<h4>Scalable, MAS-agnostic pipeline</h4>
<p>A reusable data-generation and distillation pipeline that is agnostic to the underlying MAS algorithm, enabling research on any interaction protocol.</p>
</div>
<div class="feature-card tone-warm">
<span class="feature-icon-wrap" aria-hidden="true">03</span>
<h4>Extensive empirical evaluation</h4>
<p>120 experiments across Qwen3, Gemma 3, and Llama 3 families, covering math, medical QA, long-form QA, scaling, robustness, and multimodal transfer.</p>
</div>
</div>
</div>
</section>
<!-- ============================ FRAMEWORK ============================ -->
<section class="section alt-bg" id="method">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">04 · Framework</div>
<h2 class="section-title">The AgentArk pipeline</h2>
<p class="section-lead">
Three stages: <strong>multi-agent debate</strong> → <strong>knowledge extraction</strong> → <strong>hierarchical distillation</strong>.
</p>
</div>
<div class="fig-full">
<a class="figure-link" href="./media/figures/pipeline.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: AgentArk pipeline: multi-agent debate, knowledge extraction, and three distillation strategies">
<img src="./media/figures/pipeline.png" alt="AgentArk pipeline: multi-agent debate, knowledge extraction, and three distillation strategies" width="1936" height="873" loading="lazy" decoding="async">
</a>
<p class="caption">
The pipeline proceeds through (1) multi-agent debate that produces diverse reasoning trajectories, (2) knowledge extraction that filters corrective traces with a high-capacity verifier (Qwen2.5-72B), and (3) three distillation routes: reasoning-enhanced SFT, trajectory-based data augmentation, and process-aware distillation with a PRM and GRPO.
</p>
</div>
</div>
</section>
<!-- ============================ THREE DISTILLATION STRATEGIES ============================ -->
<section class="section">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">05 · Strategies</div>
<h2 class="section-title">Three distillation strategies</h2>
</div>
<div>
<div class="methodology-step">
<h4><span class="tag-label">R-SFT</span>Reasoning-Enhanced Supervised Fine-Tuning</h4>
<p>Supervise the student on both the final consensus answer <em>and</em> the reasoning trace that produced it. A reasoning loss trains coherent intermediate rationales, while an answer loss grounds the final prediction in those rationales.</p>
</div>
<div class="methodology-step tone-teal">
<h4><span class="tag-label">DA</span>Distillation with Trajectory Augmentation</h4>
<p>For every problem, extract multiple answer-consistent but structurally <em>diverse</em> reasoning chains from the debate log. The student learns several valid paths to the same answer, improving robustness and generalization.</p>
</div>
<div class="methodology-step tone-warm">
<h4><span class="tag-label">PAD</span>Process-Aware Distillation (PRM + GRPO)</h4>
<p>Train a Process Reward Model with a contrastive step-level objective, then fine-tune the student with Group Relative Policy Optimization. This internalizes the dialectical critique-and-revision behavior of multi-agent debate within a single forward pass.</p>
</div>
</div>
</div>
</section>
<!-- ============================ MAIN RESULTS ============================ -->
<section class="section alt-bg" id="results">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">06 · Results</div>
<h2 class="section-title">Single-agent performance, multi-agent quality</h2>
<p class="section-lead">
Distilling Qwen3-32B into smaller students across GSM8K, MATH, MetaMathQA, and MedMCQA, AgentArk lifts single-agent accuracy by <strong>+4.8% on average</strong> — only marginally below a full multi-agent system, at a small fraction of the inference cost.
</p>
</div>
<div class="grid-2" style="margin-bottom: 1.5rem;">
<div class="fig-card">
<a class="figure-link" href="./media/figures/effectiveness_id-ood.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Performance on in-domain and OOD datasets">
<img src="./media/figures/effectiveness_id-ood.png" alt="Performance on in-domain and OOD datasets" width="1333" height="766" loading="lazy" decoding="async">
</a>
<div class="caption">In-domain (left) vs. OOD (right). Gains are larger ID; PAD transfers most reliably to OOD.</div>
</div>
<div class="fig-card">
<a class="figure-link" href="./media/figures/target_model_dataset.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Performance across datasets and student models">
<img src="./media/figures/target_model_dataset.png" alt="Performance across datasets and student models" width="1285" height="656" loading="lazy" decoding="async">
</a>
<div class="caption">Effects vary by dataset and student; reasoning-heavy benchmarks (MetaMathQA, GSM8K) benefit most.</div>
</div>
</div>
<div class="grid-3">
<div class="feature-card">
<span class="feature-icon-wrap"><svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="m3 12 4 4L17 6m-5 11 3 3 7-9"/></svg></span>
<h4>PAD is most consistent</h4>
<p>R-SFT and DA sometimes help but swing by dataset. PAD's step-level supervision delivers stable gains across settings.</p>
</div>
<div class="feature-card tone-teal">
<span class="feature-icon-wrap"><svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><rect x="8" y="3" width="8" height="5" rx="1"/><path d="M12 8v5M5 16v-3h14v3"/><rect x="2" y="16" width="6" height="5" rx="1"/><rect x="16" y="16" width="6" height="5" rx="1"/></svg></span>
<h4>Cross-family helps more</h4>
<p>Distilling across families (Qwen → Gemma / Llama) yields larger and more consistent gains than same-family distillation.</p>
</div>
<div class="feature-card tone-warm">
<span class="feature-icon-wrap"><svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><path d="M9 3a3 3 0 0 0-3 3v1a4 4 0 0 0-2 7 4 4 0 0 0 5 6h3V5a3 3 0 0 0-3-2Zm6 0a3 3 0 0 1 3 3v1a4 4 0 0 1 2 7 4 4 0 0 1-5 6h-3M7 11h2m6 5h2"/></svg></span>
<h4>Reasoning-heavy tasks win</h4>
<p>Biggest lifts on MetaMathQA and GSM8K; smaller on knowledge-heavy MedMCQA — distillation transfers reasoning, not facts.</p>
</div>
</div>
</div>
</section>
<!-- ============================ KEY FINDINGS ============================ -->
<section class="section" id="findings">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">07 · Findings</div>
<h2 class="section-title">Six findings from 120 experiments</h2>
</div>
<div class="findings-grid">
<div class="finding-card">
<span class="num-badge">F1</span>
<span class="finding-title">A single agent can acquire multi-agent reasoning.</span>
<p>All three reasoning-centric distillation methods boost single-agent performance, and combining them yields further gains.</p>
</div>
<div class="finding-card">
<span class="num-badge">F2</span>
<span class="finding-title">PRM capacity matters more than student size.</span>
<p>Strong PRMs lift even small students; weak PRMs limit everyone. Scaling teacher agents mostly helps larger students, with diminishing returns on small ones.</p>
</div>
<div class="finding-card">
<span class="num-badge">F3</span>
<span class="finding-title">Reasoning quality outweighs quantity.</span>
<p>Simply adding more trajectories does not improve performance; PAD's high-signal process supervision yields stable, consistent gains.</p>
</div>
<div class="finding-card">
<span class="num-badge">F4</span>
<span class="finding-title">PAD improves reasoning <em>behavior</em>, not just accuracy.</span>
<p>PAD-distilled models show better step decomposition, intermediate self-checking, and error correction than R-SFT and DA.</p>
</div>
<div class="finding-card">
<span class="num-badge">F5</span>
<span class="finding-title">Generalization & robustness improve.</span>
<p>Distilled models transfer reliably to unseen reasoning benchmarks (HotpotQA, QASPER, QMSum) and robustness evaluations such as TruthfulQA.</p>
</div>
<div class="finding-card">
<span class="num-badge">F6</span>
<span class="finding-title">AgentArk extends to multimodal LLMs.</span>
<p>Despite training on text-only reasoning data, the distilled behaviors transfer to Qwen2.5-VL, suggesting modality-agnostic reasoning transfer.</p>
</div>
</div>
</div>
</section>
<!-- ============================ SCALING & DATA DYNAMICS ============================ -->
<section class="section alt-bg" id="analysis">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">08 · Scaling</div>
<h2 class="section-title">Scaling & data dynamics</h2>
</div>
<div class="highlight-box">
<h4>Reasoning quality beats data volume.</h4>
<p>Bigger teacher ensembles mainly help larger students; for smaller ones, high-signal process supervision matters more than raw trajectory count.</p>
</div>
<h3 class="subsection-title">Scaling the number of <span class="accent">debating agents</span></h3>
<div class="grid-2">
<div class="fig-card">
<a class="figure-link" href="./media/figures/agent_scale_0.6B.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Effect of agent scale on 0.6B student">
<img src="./media/figures/agent_scale_0.6B.png" alt="Effect of agent scale on 0.6B student" width="1168" height="630" loading="lazy" decoding="async">
</a>
<div class="caption">Qwen3-0.6B saturates around 5 agents; more agents can hurt due to limited capacity.</div>
</div>
<div class="fig-card">
<a class="figure-link" href="./media/figures/agent_scale_8B.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Effect of agent scale on 8B student">
<img src="./media/figures/agent_scale_8B.png" alt="Effect of agent scale on 8B student" width="1168" height="630" loading="lazy" decoding="async">
</a>
<div class="caption">Qwen3-8B benefits from 5 → 10 → 20 agents, with diminishing returns.</div>
</div>
</div>
<h3 class="subsection-title">Data <span class="accent">quantity vs. quality</span></h3>
<div class="grid-2">
<div class="fig-card">
<a class="figure-link" href="./media/figures/data_scale_0.6B_gsm8k.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Data scaling on GSM8K">
<img src="./media/figures/data_scale_0.6B_gsm8k.png" alt="Data scaling on GSM8K" width="1000" height="600" loading="lazy" decoding="async">
</a>
<div class="caption">GSM8K: R-SFT and DA fluctuate with scale; PAD remains stable.</div>
</div>
<div class="fig-card">
<a class="figure-link" href="./media/figures/data_scale_0.6B_meta.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Data scaling on MetaMathQA">
<img src="./media/figures/data_scale_0.6B_meta.png" alt="Data scaling on MetaMathQA" width="1000" height="600" loading="lazy" decoding="async">
</a>
<div class="caption">MetaMathQA: PAD preserves high-signal trajectories instead of overwhelming the student.</div>
</div>
</div>
</div>
</section>
<!-- ============================ REASONING QUALITY ============================ -->
<section class="section">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">09 · Quality</div>
<h2 class="section-title">How distilled models actually reason</h2>
<p class="section-lead">
Reasoning-token perplexity on held-out GSM8K drops substantially after distillation, indicating more structured chains of thought. An LLM-judge evaluation (InternLM-2.5-20B-Chat) scores models on step decomposition, intermediate verification, error localization, and overall coherence.
</p>
</div>
<div>
<div class="methodology-step tone-warm">
<h4><span class="tag-label">PAD</span>Best on all four reasoning-quality dimensions</h4>
<p>PAD preserves explicit multi-step structure, self-checking behavior, and coherent reasoning flows, outperforming the base and both SFT variants.</p>
</div>
<div class="methodology-step tone-teal">
<h4><span class="tag-label">DA</span>Moderate gains on verification & error localization</h4>
<p>DA captures surface-level reasoning structure but does not fully inherit reflective self-correction behavior.</p>
</div>
<div class="methodology-step">
<h4><span class="tag-label">R-SFT</span>Smaller, less consistent improvement</h4>
<p>Direct reasoning-level supervision alone is insufficient — it improves fluency but not the dialectical dynamics of MAS.</p>
</div>
</div>
</div>
</section>
<!-- ============================ ROBUSTNESS & GENERALIZATION ============================ -->
<section class="section alt-bg">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">10 · Generalization</div>
<h2 class="section-title">Robustness & cross-domain transfer</h2>
<p class="section-lead">
Students trained only on GSM8K transfer to HotpotQA (multi-hop QA), QASPER (long-context understanding), and QMSum (summarization) — tasks far from the training distribution.
</p>
</div>
<div class="fig-full">
<a class="figure-link" href="./media/figures/generalizability_f1-score.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Open-ended generalization F1 scores on HotpotQA, QASPER, and QMSum">
<img src="./media/figures/generalizability_f1-score.png" alt="Open-ended generalization F1 scores on HotpotQA, QASPER, and QMSum" width="1795" height="684" loading="lazy" decoding="async">
</a>
<p class="caption">
F1 on three open-ended OOD benchmarks. AgentArk boosts cross-domain reasoning transfer, especially for larger students — distillation strengthens general reasoning rather than fitting dataset patterns.
</p>
</div>
<div class="highlight-box">
<h4>Distilled behaviors survive distribution shift.</h4>
<p>On TruthfulQA, all three distillation variants improve factual robustness over the base student, with PAD ranking highest — evidence that MAS distillation transfers reasoning behavior, not surface alignment.</p>
</div>
</div>
</section>
<!-- ============================ MULTIMODAL ============================ -->
<section class="section">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">11 · Multimodal</div>
<h2 class="section-title">Extension to multimodal LLMs</h2>
<p class="section-lead">
Although training data is entirely text, distilling Qwen2.5-VL-32B-Instruct into Qwen2.5-VL-3B-Instruct via the same pipeline still improves multimodal reasoning. PAD remains the strongest or near-strongest across benchmarks, suggesting AgentArk captures modality-agnostic reasoning patterns.
</p>
</div>
<div class="fig-full">
<a class="figure-link" href="./media/figures/vlm_math-gsm8k.png" target="_blank" rel="noopener" aria-label="Open full-resolution figure: Multimodal distillation results on Math and GSM8K">
<img src="./media/figures/vlm_math-gsm8k.png" alt="Multimodal distillation results on Math and GSM8K" width="993" height="358" loading="lazy" decoding="async">
</a>
<p class="caption">
Multimodal distillation on Math-derived and GSM8K-derived data, evaluated on Qwen2.5-VL-3B-Instruct. Gains are smaller than in text-only settings, as expected, but consistent.
</p>
</div>
</div>
</section>
<!-- ============================ BIBTEX ============================ -->
<section class="section bibtex-section" id="bibtex">
<div class="container">
<div class="section-head">
<div class="section-eyebrow">12 · Cite</div>
<h2 class="section-title">BibTeX</h2>
</div>
<div class="bibtex-wrap">
<div class="bibtex-bar">
<span class="file-name">agentark.bib</span>
<button type="button" class="copy-bib" data-target="#bibtex-code" aria-label="Copy BibTeX to clipboard" hidden>
<svg class="icon" viewBox="0 0 24 24" aria-hidden="true"><rect x="8" y="8" width="12" height="13" rx="2"/><path d="M16 8V3H3v13h5"/></svg><span class="copy-label">Copy</span>
</button>
</div>
<span class="sr-only copy-status" role="status" aria-live="polite"></span>
<pre id="bibtex-code"><code>@article{luo2026agentark,
title={AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent},
author={Luo, Yinyi and Jin, Yiqiao and Yu, Weichen and Zhang, Mengqi and Kumar, Srijan and Li, Xiaoxiao and Xu, Weijie and Chen, Xin and Wang, Jindong},
journal={arXiv preprint arXiv:2602.03955},
year={2026}
}</code></pre>
</div>
</div>
</section>
</main>
<!-- ============================ FOOTER ============================ -->
<footer class="site-footer">
<p>
The data, code, and model artifacts associated with this paper are intended and licensed mainly for research use.
This website is licensed under the
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">CC BY-SA 4.0 International License</a>.
</p>
<p class="footer-meta">© 2026 AIFrontierLab.</p>
</footer>
</body>
</html>