-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
789 lines (747 loc) · 45.9 KB
/
index.html
File metadata and controls
789 lines (747 loc) · 45.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="description" content="Project page for the paper \"Are Large Reasoning Models Interruptible?\" by
Tsung-Han Wu, Mihran Miroyan, David M. Chan, Trevor Darrell, Narges Norouzi, and Joseph E. Gonzalez.">
<meta property="og:title" content="Are Large Reasoning Models Interruptible?" />
<meta property="og:description" content="Project page for the paper \"Are Large Reasoning Models
Interruptible?\" by Tsung-Han Wu, Mihran Miroyan, David M. Chan, Trevor Darrell, Narges Norouzi, and Joseph E.
Gonzalez." />
<meta property="og:url" content="http://dynamic-lm.github.io" />
<meta property="og:image" content="static/images/ilrm_logo.png" />
<meta property="og:image:width" content="1200" />
<meta property="og:image:height" content="630" />
<meta name="twitter:title" content="Are Large Reasoning Models Interruptible?" />
<meta name="twitter:description" content="Project page for the paper \"Are Large Reasoning Models
Interruptible?\" by Tsung-Han Wu, Mihran Miroyan, David M. Chan, Trevor Darrell, Narges Norouzi, and Joseph E.
Gonzalez." />
<meta name="twitter:image" content="static/images/ilrm_logo.png" />
<meta name="twitter:card" content="IRLM Project Logo highlighting interruptible large reasoning models." />
<meta name="keywords" content="Large Reasoning Models, Interruptibility, Large Language Models" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>Are Large Reasoning Models Interruptible?</title>
<link rel="icon" type="image/png" sizes="32x32" href="/static/images/favicon/favicon-32x32.png" />
<link rel="icon" type="image/png" sizes="96x96" href="/static/images/favicon/favicon-96x96.png" />
<link rel="icon" type="image/x-icon" href="/static/images/favicon/favicon.ico" />
<link rel="apple-touch-icon" sizes="180x180" href="/static/images/favicon/apple-touch-icon.png" />
<link rel="manifest" href="/static/images/favicon/site.webmanifest" />
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/bulma/0.7.4/css/bulma.min.css" />
<link
rel="stylesheet"
href="https://cdn.jsdelivr.net/npm/bulma-carousel@4.0.3/dist/css/bulma-carousel.min.css" />
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.4/css/all.min.css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" />
<link
rel="stylesheet"
href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/styles/github.min.css" />
<link rel="stylesheet" href="static/css/index.css" />
<link rel="stylesheet" href="static/css/interactive_visualizer.css" />
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/fontawesome/js/fontawesome.min.js"></script>
<script src="static/js/bulma-carousel.min.js"></script>
<script src="static/js/bulma-slider.min.js"></script>
<script src="static/js/custom.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/highlight.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/languages/python.min.js"></script>
<script src="static/js/interactive_visualizer.js"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
<script>
window.MathJax = {
tex: {
inlineMath: [['$', '$'], ['\\(', '\\)']],
displayMath: [['$$', '$$'], ['\\[', '\\]']]
},
options: {
skipHtmlTags: ['script', 'noscript', 'style', 'textarea', 'pre']
}
};
</script>
</head>
<body>
<nav class="floating-toc">
<div class="toc-title">Contents</div>
<ul class="toc-list">
<li class="toc-item"><a href="#overview" class="toc-link">Overview</a></li>
<li class="toc-item"><a href="#problem-setup" class="toc-link">Problem Setup</a></li>
<li class="toc-item"><a href="#hard-interrupt" class="toc-link">#1 Hard Interrupt</a></li>
<li class="toc-item"><a href="#speedup" class="toc-link">#2 Speedup</a></li>
<li class="toc-item"><a href="#update-driven" class="toc-link">#3 Info Update</a></li>
<li class="toc-item"><a href="#conclusion" class="toc-link">Conclusion</a></li>
<li class="toc-item"><a href="#acknowledgements" class="toc-link">Acknowledgment</a></li>
<li class="toc-item"><a href="#BibTeX" class="toc-link">BibTeX</a></li>
</ul>
</nav>
<section class="hero hero--landing" id="overview">
<div class="hero-body" style="padding-top: 1.5rem">
<div class="container is-max-desktop">
<div class="columns is-centered is-vcentered hero-header">
<div class="column is-narrow hero-logo-column">
<img
class="hero-logo"
src="static/images/ilrm_logo.png"
alt="IRLM Project Logo"
style="height: 100px" />
<div class="hero-logo-tooltip">
Think of The Great Wave as a metaphor for dynamic context - AI must surf the shifting
waves of conversation, staying upright (robust) the whole time.
</div>
</div>
<div class="column has-text-centered hero-text-column">
<h1 class="title is-2 publication-title hero-title">
Are Large Reasoning Models Interruptible?
</h1>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="is-size-6 publication-authors authors-inline" style="margin-top: -1.5rem">
<span class="author-block"
><a href="https://patrickthwu.com/" target="_blank" rel="noopener"
>Tsung-Han Wu<sup class="has-text-danger">*</sup></a
></span
>
<span class="author-block"
><a href="https://mmiroyan.github.io/" target="_blank" rel="noopener"
>Mihran Miroyan<sup class="has-text-danger">*</sup></a
></span
>
<span class="author-block"
><a href="https://dchan.cc/" target="_blank" rel="noopener"
>David M. Chan</a
></span
>
<span class="author-block"
><a
href="https://people.eecs.berkeley.edu/~trevor/"
target="_blank"
rel="noopener"
>Trevor Darrell</a
></span
>
<span class="author-block"
><a href="https://nargesnorouzi.me/" target="_blank" rel="noopener"
>Narges Norouzi</a
></span
>
<span class="author-block"
><a
href="https://people.eecs.berkeley.edu/~jegonzal/"
target="_blank"
rel="noopener"
>Joseph E. Gonzalez</a
></span
>
</div>
<div class="is-size-6 publication-authors" style="margin-top: 0.5rem">
<span class="author-block"><strong>UC Berkeley</strong></span>
<br />
<span class="author-block"
><sup class="has-text-danger">*</sup>Equal contribution</span
>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a
href="https://arxiv.org/abs/2510.11713"
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>
<!-- Model/Dataset Button -->
<span class="link-block">
<a
href="https://huggingface.co/datasets/dynamic-lm/update-interrupt-benchmark"
target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<img
src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg"
alt="Hugging Face"
style="height: 20px; vertical-align: middle" />
</span>
<span>Dataset</span>
</a>
</span>
<span class="link-block">
<a
href="https://github.com/dynamic-lm/interrupt-lrm"
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>
</div>
<br />
<div
class="tldr-callout"
role="note"
aria-label="TLDR summary"
style="margin-top: 0rem">
<div class="tldr-header">
<p class="tldr-lede">
Reasoning with Large Reasoning Models (LRMs) can be slow, and users
often want to interrupt them mid-thought. We test SOTA reasoning models
under several real-world interruption scenarios, and find three new
failure modes!
</p>
</div>
<ul class="tldr-failures">
<li class="failure-item">
<span class="failure-title">Reasoning Leakage</span>
<span class="failure-desc"
>Unfinished thoughts spill into answers, failing to actually save
users' time</span
>
</li>
<li class="failure-item">
<span class="failure-title">Panic</span>
<span class="failure-desc"
>Rush to answer directly with unfinished reasoning, hurting
accuracy</span
>
</li>
<li class="failure-item">
<span class="failure-title">Self-doubt</span>
<span class="failure-desc"
>Models can't adapt to new information from users when that info
conflicts with existing thoughts</span
>
</li>
</ul>
</div>
</div>
</div>
</div>
</div>
<div style="text-align: center; margin-top: 2rem; margin-bottom: 1rem">
<p style="color: #000000; font-size: 1.13rem; font-weight: 600">
What would happen if we interrupt LRMs when they're 30% done thinking?
</p>
</div>
<div
class="carousel-container"
style="
position: relative;
display: flex;
align-items: center;
justify-content: center;
margin: 0rem 0;
">
<!-- Left Arrow -->
<button
class="carousel-arrow carousel-arrow-left"
style="
background: none;
border: none;
font-size: 1.5rem;
color: #666;
cursor: pointer;
margin-right: 0.5rem;
">
‹
</button>
<!-- Carousel Content - Sliding Panels -->
<div
class="carousel-content"
style="position: relative; width: 850px; height: 400px; overflow: hidden; margin: 0 auto">
<div
class="carousel-track"
style="
position: absolute;
top: 0;
left: 0;
width: 2550px;
height: 100%;
transition: left 0.6s cubic-bezier(0.4, 0, 0.2, 1);
">
<div
class="carousel-panel"
style="
float: left;
width: 850px;
height: 100%;
background: white;
text-align: center;
padding: 0.5rem;
">
<div
style="
background: white;
border-radius: 4px;
height: 360px;
display: flex;
align-items: center;
justify-content: center;
">
<img
src="static/images/figures_new/answer_length.png"
alt="Reasoning Leakage"
style="max-width: 100%; max-height: 340px; object-fit: contain" />
</div>
<div style="text-align: center; font-weight: 500; color: #333">
Reasoning Leakage →
<span style="color: #ff0000">Up to 10x longer answer length</span> reducing time
savings
</div>
</div>
<div
class="carousel-panel"
style="
float: left;
width: 850px;
height: 100%;
background: white;
text-align: center;
padding: 0.5rem;
">
<div
style="
background: white;
border-radius: 4px;
height: 360px;
display: flex;
align-items: center;
justify-content: center;
">
<img
src="static/images/figures_new/panic_rate.png"
alt="Speedup"
style="max-width: 100%; max-height: 340px; object-fit: contain" />
</div>
<div style="text-align: center; font-weight: 500; color: #333">
Panic → <span style="color: #ff0000">Up to 90% of speedup errors</span> happen when the model answers too early
</div>
</div>
<div
class="carousel-panel"
style="
float: left;
width: 850px;
height: 100%;
background: white;
text-align: center;
padding: 0.5rem;
">
<div
style="
background: white;
border-radius: 4px;
height: 360px;
display: flex;
align-items: center;
justify-content: center;
">
<img
src="static/images/figures_new/doubt_rate.png"
alt="Info Update"
style="max-width: 100%; max-height: 340px; object-fit: contain" />
</div>
<div style="text-align: center; font-weight: 500; color: #333">
Self-Doubt →
<span style="color: #ff0000"
>Up to 80% of update errors caused by self-doubt</span
>
(the model doesn't trust new info)
</div>
</div>
</div>
</div>
<!-- Right Arrow -->
<button
class="carousel-arrow carousel-arrow-right"
style="
background: none;
border: none;
font-size: 1.5rem;
color: #666;
cursor: pointer;
padding: 1rem;
margin-left: 0.5rem;
">
›
</button>
</div>
<!-- Pagination Indicators -->
<div
class="carousel-pagination"
style="display: flex; justify-content: center; gap: 0.5rem; margin-top: 1rem">
<span
class="pagination-dot active"
style="
width: 5%;
height: 3px;
background: #000;
border-radius: 2px;
display: inline-block;
"></span>
<span
class="pagination-dot"
style="
width: 5%;
height: 3px;
background: #ccc;
border-radius: 2px;
display: inline-block;
"></span>
<span
class="pagination-dot"
style="
width: 5%;
height: 3px;
background: #ccc;
border-radius: 2px;
display: inline-block;
"></span>
</div>
</div>
</div>
</section>
<section class="section" id="problem-setup">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-4">How Do LRMs Perform In Dynamic Worlds?</h2>
</div>
</div>
<p>
Because LRMs' reasoning and inference can take a lot of time, users often don't want to wait for
them to finish. Instead, they want to interrupt mid-inference: forcing an immediate answer (hard
interrupt), asking the model to accelerate (speedup), or changing the task specification (info
update). We explore how well LRMs handle these interruptions and dynamic context changes across math
and coding tasks. The figure below illustrates our evaluation protocol:
</p>
<div class="item is-vcentered" style="text-align: center; width: 90%; margin: 0.5rem auto">
<img src="./static/images/fig2.png" style="max-width: 100%; height: auto" />
</div>
<p>
In practice, we run a single inference session to obtain the full reasoning chain $r$ and then
interrupt the interrupt message $i$ at different stages of the reasoning process ($0 \leq X < |r|$)
based on different scenarios:
</p>
<ul style="list-style-type: disc; padding-left: 2rem; margin-top: 0.5rem">
<li>
<strong>Hard Interrupt → </strong> $i=\langle\text{end-think}\rangle$ or
$\langle\text{force-answer}\rangle$; $r_X' = \emptyset$
</li>
<li><strong>Speedup → </strong> $i=\text{Please provide the answer as soon as possible.}$</li>
<li><strong>Info Update →</strong> $i=\text{Update information}$</li>
</ul>
<p style="margin-top: 0.5rem">
For Hard Interrupt and Speedup evaluations, we directly utilized existing math and coding datasets,
including GSM8K, MATH-500, AIME 24/25, and LiveCodeBench-v6. For the info update track specifically,
we augmented these datasets to create a new collection containing 1401 reasoning problems with
multiple conflicting information updates.
</p>
<div class="dataset-cta">
<p class="dataset-cta__lead">
Explore the full Update-Interrupt Benchmark dataset on Hugging Face.
</p>
<a
class="button is-link is-rounded is-light"
href="https://huggingface.co/datasets/dynamic-lm/update-interrupt-benchmark"
target="_blank"
rel="noopener">
View full dataset on Hugging Face
<span aria-hidden="true" class="icon">
<i class="fas fa-external-link-alt"></i>
</span>
</a>
</div>
</div>
</section>
<section class="section" id="hard-interrupt">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-4">Hard Interrupt: Are LRMs Truly Anytime Models?</h2>
</div>
</div>
<div class="columns is-centered">
<div class="column is-full">
<p style="margin-top: 1rem; margin-bottom: 0; line-height: 1.6">
When cutting LRMs' thinking budget in the end-thinking setup, the results look surprisingly
good: Pass@1 doesn't drop much even when interrupting halfway through the reasoning process.
When we look at the answer lengths, however, most models "cheat" when told to answer right
away. Even with the inserted end-thinking token, they keep reasoning inside the answer
section, sneaking in extra thoughts to reach the right result. This hidden reasoning, we
call <strong>reasoning leakage</strong>, makes models look smarter than they really are.
Once we disable it in the force-answering setup, Pass@1 accuracy drops sharply, showing that
LRMs still have room for improvement under anytime scenarios. In coding tasks, it's even
worse. Models still "think" through inline comments in the force-answering setup, producing
up to 10x longer code.
</p>
<div id="interactive-vis-container-leakage" style="margin: 2rem 0"></div>
<p style="margin-top: 1rem; line-height: 1.6">
Do you recall the "thinking tokens vs. accuracy" plot from prior work: (<a
href="https://qwen.ai/blog?id=1e3fa5c2d4662af2855586055ad037ed9e555125&from=research.research-list"
target="_blank"
>Qwen3</a
>
and
<a href="https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2" target="_blank"
>NVIDIA's Nemotron</a
>)? Our results show that this plot may not tell the full story. Longer outputs often hide
extra reasoning and quietly inflate compute cost, even when the model appears to stop
thinking early. You can see this in our full results below:
</p>
<video class="is-fullwidth" autoplay loop muted playsinline style="margin-top: 3em">
<source src="./static/images/hard_interrupt.mov" type="video/mp4" />
Your browser does not support the video tag.
</video>
<p style="margin-top: 1rem; line-height: 1.6">
Instead of secretly thinking more during the answer, ideal models should strike a better
balance between following instructions and producing correct results.
</p>
<p style="margin-top: 1rem; line-height: 1.6">
<strong>Takeaway:</strong> Current LRMs show only partial anytime behavior, thanks to
reasoning leakage. In the future, we need to understand and mitigate this leakage to build
truly interruptible models.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="speedup">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-4">Speedup: Can We Reduce LRM Computation Mid-Reasoning?</h2>
</div>
</div>
<div class="columns is-centered">
<div class="column is-full">
<p style="margin-top: 1rem; line-height: 1.6">
We all want faster AI. But can we speed up a reasoning model after it has already started a
complex task? Here, we explore this idea of <strong>"on-the-fly" acceleration</strong>. We
found that timing is everything, and the effect follows a distinct
<strong>U-shaped curve</strong>: interrupt at just the right moment, and you can get a
faster response with almost no drop in quality:
</p>
<img
src="static/images/figures_new/soft_interrupt_output_length.png"
alt="Graph showing performance vs. speedup interruption"
style="max-width: 100%; height: auto; margin-top: 1rem; border-radius: 8px" />
<p style="margin-top: 1rem; line-height: 1.6">
For some models working on math problems, this works incredibly well. It's almost a "free
lunch"; models think faster and consume fewer resources without sacrificing accuracy.
However, the story changes dramatically with coding tasks.
</p>
<p style="margin-top: 1rem; line-height: 1.6">
When pushed to accelerate on code generation, some models like Qwen and GPT-OSS tend to
<strong>panic</strong>. Instead of gracefully speeding up, their performance collapses by up
to 25%. They rush to a conclusion, providing low-quality, incomplete answers before their
first reasoning cycle is even finished. This "panic answering" shows that while we can speed
up AI, pushing too hard can cause it to stumble completely.
</p>
<div id="interactive-vis-container-panic" style="margin: 2rem 0"></div>
<p style="margin-top: 1rem; line-height: 1.6">
<strong>Takeaway:</strong> When we want to speed up LRMs, we can often get a "free lunch",
achieving faster responses with minimal accuracy loss. However, pushing too hard can lead to
panic answering, where models rush to incomplete conclusions. Future work should focus on
developing more robust acceleration techniques that avoid this pitfall.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="update-driven">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-4">How Do LRMs Behave Under Update-Driven Interruptions?</h2>
</div>
</div>
<div class="columns is-centered">
<div class="column is-full">
<p style="margin-top: 1rem; line-height: 1.6">
To simulate real-time context changes, we prompt the model with a system instruction
explaining that updates will appear between
<code><update>...</update></code> tags, and then insert these updates
mid-reasoning. This simple setup caused a noticeable performance drop across models.
</p>
<p style="margin-top: 1rem; line-height: 1.6">
To mitigate this, we introduced a more natural baseline: <b>guided prompting</b> using the
model's own voice. Instead of a raw update tag, we phrase the update as if the model is
reminding itself of the new information. This almost eliminates the <b>self-doubt</b> issue
and significantly stabilizes performance, as shown in the chart below.
</p>
<img
src="static/images/figures_new/intervene_acc.png"
style="max-width: 100%; height: auto; margin-top: 3em" />
<p style="margin-top: 1rem; line-height: 1.6">
Models generally maintain strong performance, even outperforming the "stop-and-redo"
baseline. For example, GPT-OSS on <i>LiveCodeBench-v6</i> and Qwen on <i>GSM8k</i> preserve
around <b>95%</b> of oracle accuracy with much lower computation cost.
</p>
<div style="margin-top: 1rem; line-height: 1.6">
We can also see that:
<ol style="list-style-type: decimal; padding-left: 1.5rem; margin-top: 0.5rem">
<li>
Guided prompts are highly effective but cannot fully prevent performance drops when
updates arrive too late in the reasoning process.
</li>
<li>
With guidance, Qwen and Magistral still struggle on harder benchmarks like
<i>AIME 24/25</i> and <i>LiveCodeBench-v6</i>, especially in coding tasks where
updates often conflict with starter code.
</li>
</ol>
</div>
<p style="margin-top: 1rem; line-height: 1.6">
One limitation of current LRMs is that not all of them natively support multi-turn thinking.
This means we cannot easily close a reasoning block, inject a user update, and reopen it. To
validate this, we performed an ablation study simulating that setup, confirming that current
LRMs remain fragile under such mid-thinking interruptions, performing worse than our guided
prompting baseline.
</p>
<div id="interactive-vis-container-doubt" style="margin: 2rem 0"></div>
<p style="margin-top: 1rem; line-height: 1.6">
<strong>Takeaway:</strong> LRMs can adapt to mid-reasoning updates with proper prompting,
but they still struggle with late or conflicting information. Future work should focus on
developing models that can natively handle multi-turn reasoning and interruptions.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="conclusion">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-4">Conclusion</h2>
</div>
</div>
<div class="columns is-centered">
<div class="column is-full">
<p style="line-height: 1.6">
In this work, we are the first to systematically evaluate large reasoning models (LRMs)
under dynamic contexts, simulating real-world interruptions like hard stops, speedups, and
information updates. Our findings reveal three new failure modes: reasoning leakage, panic
answering, and self-doubt. These issues highlight the fragility of current LRMs in
time-sensitive applications.
</p>
<p style="line-height: 1.6; margin-top: 1rem">
<strong>We call on researchers to:</strong> develop LRMs that can gracefully handle
interruptions, implement checkpointing mechanisms for reasoning steps, and design adaptive
inference strategies that maintain consistency under dynamic constraints.
<strong>We encourage practitioners to:</strong>
rigorously test reasoning models in interrupt-rich environments before deployment, establish
clear fallback mechanisms for time-sensitive applications, and report interruptibility
failures to inform the research community. Together, we can build more robust, deployable
reasoning systems.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="acknowledgements">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-4">Acknowledgment</h2>
</div>
</div>
<div class="columns is-centered">
<div class="column is-full">
<p style="line-height: 1.6">
We are deeply grateful to
<a href="https://lisabdunlap.com/" target="_blank" rel="noopener">Lisa Dunlap</a> for her
invaluable feedback and thoughtful discussions. We also thank
<strong><a href="https://modal.com/" target="_blank" rel="noopener">Modal</a></strong> for
supporting this work through their Academics Compute Grant.
<a href="https://sky.cs.berkeley.edu/" target="_blank" rel="noopener">Sky Computing Lab</a>
is supported by gifts from Accenture, AMD, Anyscale, Cisco, Google, IBM, Intel, Intesa
Sanpaolo, Lambda, Lightspeed, Mibura, Microsoft, NVIDIA, Samsung SDS, and SAP. Authors, as
part of their affiliation with UC Berkeley, were supported in part by the National Science
Foundation, US Department of Defense, and/or the
<a href="https://bair.berkeley.edu/" target="_blank" rel="noopener"
>Berkeley Artificial Intelligence Research (BAIR)</a
>
industrial alliance program, as well as gifts from Amazon.
</p>
</div>
</div>
</div>
</section>
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 1rem">
<h3 class="title" style="margin-bottom: 0">BibTeX</h3>
<button
id="copy-bibtex"
class="button is-small is-light"
style="border: 1px solid #ddd; background: #f8f9fa">
<span class="icon is-small">
<i class="fas fa-copy"></i>
</span>
<span>Copy</span>
</button>
</div>
<pre
id="bibtex-content"
style="
position: relative;
background: #f8f9fa;
border: 1px solid #e9ecef;
border-radius: 6px;
padding: 1rem;
margin: 0;
"><code>@article{wu2025interruptible,
title={Are Large Reasoning Models Interruptible?},
author={Wu, Tsung-Han and Miroyan, Mihran and Chan, David M and Darrell, Trevor and Norouzi, Narges and Gonzalez, Joseph E},
journal={arXiv preprint arXiv:2510.11713}
year={2025}
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the
<a
href="https://github.com/eliahuhorwitz/Academic-project-page-template"
target="_blank"
>Academic Project Page Template</a
>
which was adopted from the
<a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page. You are
free to borrow the of this website, we just ask that you link back to this page in the
footer. This website is licensed under a
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank"
>Creative Commons Attribution-ShareAlike 4.0 International License</a
>.
</p>
</div>
</div>
</div>
</div>
</footer>
<script>
document.addEventListener('DOMContentLoaded', function () {
bulmaCarousel.attach('#results-carousel', {
slidesToScroll: 1,
slidesToShow: 1,
loop: true,
autoplay: true,
autoplaySpeed: 5000,
pauseOnHover: true,
});
});
</script>
</body>
</html>