-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
627 lines (579 loc) · 29.2 KB
/
index.html
File metadata and controls
627 lines (579 loc) · 29.2 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
---
title: Home
---
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
{% seo %}
<link rel="icon" href="assets/images/favicon.svg" type="image/svg+xml">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Noto+Sans:wght@300;400;600;700&display=swap" rel="stylesheet">
<link rel="stylesheet" href="assets/css/style.css?v=1.27">
</head>
<body>
<div class="page-container">
<!-- Title Section -->
<header class="paper-header" id="top">
<div class="logo" style="margin-bottom: 1.5rem;">
<a href="https://cvpr.thecvf.com/" target="_blank">
<img src="assets/images/cvpr-navbar-logo.svg" alt="CVPR 2026 Logo" style="height: 60px;">
</a>
</div>
<h1 class="paper-title">HERBench: A Benchmark for Multi-Evidence Integration in Video Question Answering</h1>
<div class="authors">
<span class="author">Dan Ben-Ami<sup>1,*</sup>, Gabriele Serussi<sup>1,*</sup>, Kobi Cohen<sup>2</sup>, Chaim
Baskin<sup>1</sup></span>
</div>
<div class="affiliations">
<span class="affiliation"><sup>1</sup>INSIGHT Lab,
Ben-Gurion University of the Negev, Israel</span>
<span class="affiliation"><sup>2</sup>Ben-Gurion University of the Negev, Israel</span>
</div>
<div class="venue" style="font-size: 1.5rem; font-weight: 700; margin-top: 0.5rem; margin-bottom: 1rem;">CVPR 2026</div>
<div class="publication-logos">
<a href="https://insight.bgu.ac.il/" target="_blank">
<img src="assets/images/insight_logo.png" alt="INSIGHT Lab">
</a>
<a href="https://www.bgu.ac.il/" target="_blank">
<img src="assets/images/bgu_logo.png" alt="Ben-Gurion University of the Negev">
</a>
</div>
<div class="contribution-note">
<span><sup>*</sup> Equal contribution</span>
</div>
<p class="equal-contrib">VideoQA benchmark where every question requires ≥3 dispersed evidence segments</p>
<div class="paper-links">
<a href="https://github.com/DanBenAmi/HERBench/blob/main/USAGE_GUIDE.md" class="paper-btn paper-btn-primary"
target="_blank" rel="noopener">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" viewBox="0 0 6.35 6.35">
<g id="layer1" style="display:inline">
<path
d="M 4.8168655,0.00127766 C 4.6890373,0.00562477 4.5613573,0.02597387 4.4370455,0.06070573 3.9398008,0.19964558 3.4688398,0.53025887 3.0211092,0.92421897 2.1256507,1.712159 1.3105118,2.7893392 0.8806723,3.2191694 a 0.26460945,0.26460945 0 0 0 0,0.374654 L 2.7849474,5.4981 a 0.26460945,0.26460945 0 0 0 0.3731049,0 C 3.590482,5.06566 4.6594992,4.2400777 5.4374986,3.3344079 5.8265128,2.8815777 6.1505428,2.4049009 6.2849934,1.9045209 6.4194441,1.4041409 6.3263981,0.85817657 5.9041389,0.4358767 5.5874195,0.1191588 5.2003552,-0.01176471 4.8168655,0.00127766 Z M 4.4845858,1.0885499 c 0.2062348,3.439e-4 0.4126019,0.07877 0.5684415,0.234611 0.3116818,0.3116808 0.3116791,0.8236527 0,1.1353314 -0.3116845,0.3116855 -0.8251984,0.3127142 -1.1368829,0.00103 -0.3116792,-0.3116813 -0.3116845,-0.8262295 0,-1.1379151 C 4.0719866,1.1657698 4.2783537,1.0882027 4.4845858,1.0885483 Z M 2.1395096,2.6026696 3.7699012,4.2330609 C 3.6323708,4.3548619 3.4989362,4.4745025 3.3735422,4.5849774 L 1.7901776,3.0036787 C 1.9002601,2.8770351 2.0184998,2.7418139 2.1395096,2.6026696 Z"
id="path660" style="fill:currentColor" />
<path
d="M 1.562284,1.1097372 A 0.26460945,0.26460945 0 0 0 1.3824494,1.1857017 L 0.1034588,2.4667605 a 0.26460945,0.26460945 0 0 0 0.1658805,0.451652 l 0.5307171,0.041858 C 1.2014267,2.5085594 1.746389,1.8496148 2.3720545,1.2156739 L 1.6051756,1.1112876 a 0.26460945,0.26460945 0 0 0 -0.042892,-0.00156 z M 1.9426226,2.5773481 C 1.8103071,2.7294949 1.6817937,2.876367 1.5607335,3.0155643 A 0.26460945,0.26460945 0 0 0 1.7359168,2.9220299 L 1.9875806,2.6223064 Z m 3.2142748,1.407666 C 4.5262947,4.6218117 3.8676357,5.1766596 3.4154019,5.5828493 l 0.041341,0.5229656 A 0.26460945,0.26460945 0 0 0 3.9078788,6.2716962 L 5.1894517,4.9921878 A 0.26460945,0.26460945 0 0 0 5.2633498,4.7699791 Z M 3.7523303,4.3870566 3.4531236,4.6387209 A 0.26460945,0.26460945 0 0 0 3.3595907,4.8201052 C 3.4985102,4.6975791 3.6454942,4.5664428 3.7972909,4.4320152 Z"
id="path666" style="fill:currentColor" />
<path
d="m 1.3239925,4.7619079 a 0.2645835,0.2645835 0 0 0 -0.17968,0.0762 l -1.0586,1.05859 a 0.2645835,0.2645835 0 0 0 0,0.375 0.2645835,0.2645835 0 0 0 0.375,0 l 1.0586,-1.05859 a 0.2645835,0.2645835 0 0 0 0,-0.375 0.2645835,0.2645835 0 0 0 -0.19532,-0.0762 z"
id="path672" style="fill:currentColor" />
<path
d="m 0.7947025,3.9669879 a 0.2645835,0.2645835 0 0 0 -0.17969,0.0781 l -0.5293,0.52929 a 0.2645835,0.2645835 0 0 0 0,0.375 0.2645835,0.2645835 0 0 0 0.375,0 l 0.5293,-0.52929 a 0.2645835,0.2645835 0 0 0 0,-0.375 0.2645835,0.2645835 0 0 0 -0.19531,-0.0781 z"
id="path674" style="fill:currentColor" />
<path
d="m 2.1169625,5.2912079 a 0.2645835,0.2645835 0 0 0 -0.17773,0.0762 l -0.5293,0.5293 a 0.2645835,0.2645835 0 0 0 0,0.375 0.2645835,0.2645835 0 0 0 0.37305,0 l 0.52929,-0.5293 a 0.2645835,0.2645835 0 0 0 0,-0.375 0.2645835,0.2645835 0 0 0 -0.19531,-0.0762 z"
id="path676" style="fill:currentColor" />
</g>
</svg>
Get Started
</a>
<a href="https://huggingface.co/datasets/DanBenAmi/HERBench" class="paper-btn paper-btn-primary" target="_blank"
rel="noopener">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" viewBox="0 0 16 16">
<path
d="M4.318 2.687C5.234 2.271 6.536 2 8 2s2.766.27 3.682.687C12.644 3.125 13 3.627 13 4c0 .374-.356.875-1.318 1.313C10.766 5.729 9.464 6 8 6s-2.766-.271-3.682-.687C3.356 4.875 3 4.373 3 4c0-.373.356-.875 1.318-1.313ZM13 5.698V7c0 .374-.356.875-1.318 1.313C10.766 8.729 9.464 9 8 9s-2.766-.271-3.682-.687C3.356 7.875 3 7.373 3 7V5.698c.271.202.58.378.904.525C4.978 6.711 6.427 7 8 7s3.022-.289 4.096-.777A4.92 4.92 0 0 0 13 5.698ZM14 4c0-1.007-.875-1.755-1.904-2.223C11.022 1.289 9.573 1 8 1s-3.022.289-4.096.777C2.875 2.245 2 2.993 2 4v9c0 1.007.875 1.755 1.904 2.223C4.978 15.71 6.427 16 8 16s3.022-.289 4.096-.777C13.125 14.755 14 14.007 14 13V4Zm-1 4.698V10c0 .374-.356.875-1.318 1.313C10.766 11.729 9.464 12 8 12s-2.766-.271-3.682-.687C3.356 10.875 3 10.373 3 10V8.698c.271.202.58.378.904.525C4.978 9.71 6.427 10 8 10s3.022-.289 4.096-.777A4.92 4.92 0 0 0 13 8.698Zm0 3V13c0 .374-.356.875-1.318 1.313C10.766 14.729 9.464 15 8 15s-2.766-.271-3.682-.687C3.356 13.875 3 13.373 3 13v-1.302c.271.202.58.378.904.525C4.978 12.71 6.427 13 8 13s3.022-.289 4.096-.777A4.92 4.92 0 0 0 13 11.698Z" />
</svg>
Dataset
</a>
<a href="https://github.com/DanBenAmi/HERBench" class="paper-btn paper-btn-primary" target="_blank"
rel="noopener">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" viewBox="0 0 16 16">
<path
d="M8 0C3.58 0 0 3.58 0 8c0 3.54 2.29 6.53 5.47 7.59.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.012 8.012 0 0 0 16 8c0-4.42-3.58-8-8-8z" />
</svg>
Code
</a>
<a href="https://arxiv.org/abs/2512.14870" class="paper-btn paper-btn-primary" target="_blank" rel="noopener">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" viewBox="0 0 16 16">
<path
d="M14 14V4.5L9.5 0H4a2 2 0 0 0-2 2v12a2 2 0 0 0 2 2h8a2 2 0 0 0 2-2zM9.5 3A1.5 1.5 0 0 0 11 4.5h2V14a1 1 0 0 1-1 1H4a1 1 0 0 1-1-1V2a1 1 0 0 1 1-1h5.5v2z" />
</svg>
arXiv
</a>
</div>
</header>
<!-- Teaser Section -->
<section class="teaser-section">
<div class="teaser-video-container">
<video class="teaser-video" autoplay muted loop playsinline controls>
<source src="assets/images/video_teaser.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</section>
<!-- Abstract -->
<section class="abstract-section" id="abstract">
<h2>Abstract</h2>
<p class="abstract-text">
Video Large Language Models (Video-LLMs) are rapidly improving, yet current Video Question Answering
(VideoQA) benchmarks often allow questions to be answered from a single salient cue, under-testing reasoning
that
must aggregate multiple, temporally separated visual evidence. In this direction, we present
<strong>HERBench</strong>, a VideoQA
benchmark purpose-built to assess multi-evidence integration across time. Each question is constructed to
require
aggregating at least three non-overlapping evidential cues across distinct video segments (so neither language
priors
nor a single snapshot can suffice). HERBench comprises 26K five-way multiple-choice questions
organized into
twelve
compositional tasks that probe identity binding, cross-entity relations, temporal ordering, co-occurrence
verification,
and counting. To make evidential demand measurable, we introduce the <em>Minimum Required Frame-Set</em>
(MRFS)-the
smallest number of frames a model must fuse to answer correctly-and show that HERBench imposes
substantially
higher demand than prior datasets (mean MRFS 5.5 vs. 2.6-4.2). Evaluating 13 state-of-the-art Video-LLMs on
HERBench reveals pervasive failures: accuracies of 31-42% are only slightly above the 20%
random-guess baseline.
We disentangle this failure into two critical bottlenecks: (1) a <strong>retrieval deficit</strong>, where frame
selectors
overlook key
evidence, and (2) a <strong>fusion deficit</strong>, where models fail to integrate information even when all
necessary evidence
is
provided. By making cross-time evidence both unavoidable and quantifiable, HERBench establishes
a principled
target
for advancing robust, compositional video understanding.
</p>
</section>
<!-- Benchmark Overview Figure -->
<section class="method-section" id="overview">
<h2>Benchmark Overview</h2>
<div class="method-figure">
<img src="assets/images/Teaser_plot.png" alt="HERBench Overview" class="method-img">
</div>
<div class="method-description">
<p>
HERBench enforces high evidential requirements by design: questions draw on dispersed cues across long-form
videos, and answers are balanced to prevent positional bias. The MRFS (Minimum Required Frame Set) metric
reports the smallest number of frames a model with a fixed selector must fuse to answer correctly,
separating genuine multi-frame reasoning from single-cue shortcuts.
</p>
</div>
</section>
<!-- Leaderboard -->
<section class="results-section" id="leaderboard">
<h2>Leaderboard</h2>
<p class="section-description">Top-1 accuracy (%) with a 16-frame uniform budget.</p>
<div class="table-wrapper">
<table>
<thead>
<tr>
<th>Rank</th>
<th>Model</th>
<th>Selector</th>
<th>Frames</th>
<th>Overall</th>
<th>TR&C</th>
<th>R&T</th>
<th>GC&V</th>
<th>ME&N</th>
</tr>
</thead>
<tbody>
<tr>
<td>1</td>
<td>Ovis-2.5-9B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">42.1</td>
<td style="text-align:left">18.9</td>
<td style="text-align:left">73.5</td>
<td style="text-align:left">46.8</td>
<td style="text-align:left">29.2</td>
</tr>
<tr>
<td>2</td>
<td>InternVL3.5-14B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">41.5</td>
<td style="text-align:left">37.7</td>
<td style="text-align:left">69.3</td>
<td style="text-align:left">31.1</td>
<td style="text-align:left">27.8</td>
</tr>
<tr>
<td>3</td>
<td>InternVL3.5-8B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">41.1</td>
<td style="text-align:left">33.6</td>
<td style="text-align:left">70.2</td>
<td style="text-align:left">29.7</td>
<td style="text-align:left">30.8</td>
</tr>
<tr>
<td>4</td>
<td>Gemini-2.5-Flash</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">40.3</td>
<td style="text-align:left">29.7</td>
<td style="text-align:left">69.9</td>
<td style="text-align:left">34.9</td>
<td style="text-align:left">26.8</td>
</tr>
<tr>
<td>5</td>
<td>MiniCPM-V4.5-8B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">39.9</td>
<td style="text-align:left">23.8</td>
<td style="text-align:left">71.1</td>
<td style="text-align:left">39.7</td>
<td style="text-align:left">24.9</td>
</tr>
<tr class="hidden-row">
<td>6</td>
<td>Qwen2.5-VL-72B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">39.7</td>
<td style="text-align:left">26.9</td>
<td style="text-align:left">70.9</td>
<td style="text-align:left">36.6</td>
<td style="text-align:left">24.4</td>
</tr>
<tr class="hidden-row">
<td>7</td>
<td>GPT-4.1</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">39.4</td>
<td style="text-align:left">25.4</td>
<td style="text-align:left">66.0</td>
<td style="text-align:left">37.1</td>
<td style="text-align:left">29.0</td>
</tr>
<tr class="hidden-row">
<td>8</td>
<td>Qwen3-VL-8B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">38.3</td>
<td style="text-align:left">19.0</td>
<td style="text-align:left">68.7</td>
<td style="text-align:left">40.6</td>
<td style="text-align:left">25.2</td>
</tr>
<tr class="hidden-row">
<td>9</td>
<td>LLaVA-OneVision1.5-8B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">38.1</td>
<td style="text-align:left">26.1</td>
<td style="text-align:left">67.7</td>
<td style="text-align:left">33.6</td>
<td style="text-align:left">24.9</td>
</tr>
<tr class="hidden-row">
<td>10</td>
<td>Qwen2.5-VL-7B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">35.9</td>
<td style="text-align:left">21.8</td>
<td style="text-align:left">60.6</td>
<td style="text-align:left">38.7</td>
<td style="text-align:left">22.6</td>
</tr>
<tr class="hidden-row">
<td>11</td>
<td>LLaVA-OneVision-7B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">35.6</td>
<td style="text-align:left">27.3</td>
<td style="text-align:left">59.1</td>
<td style="text-align:left">30.1</td>
<td style="text-align:left">26.0</td>
</tr>
<tr class="hidden-row">
<td>12</td>
<td>Gemma-3-27B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">33.8</td>
<td style="text-align:left">32.0</td>
<td style="text-align:left">58.4</td>
<td style="text-align:left">21.5</td>
<td style="text-align:left">23.5</td>
</tr>
<tr class="hidden-row">
<td>13</td>
<td>LLaMA-4-Scout-17B</td>
<td>Uniform</td>
<td>16</td>
<td style="text-align:left">31.4</td>
<td style="text-align:left">18.8</td>
<td style="text-align:left">57.3</td>
<td style="text-align:left">25.5</td>
<td style="text-align:left">24.2</td>
</tr>
<tr>
<td colspan="9" class="baseline-row">Random baseline (5-way MCQ): 20%</td>
</tr>
</tbody>
</table>
<button id="toggle-leaderboard" class="show-more-btn">Show More</button>
</div>
<div class="highlight-card">
<h3>Observed bottlenecks</h3>
<p><strong>Evidence retrieval:</strong> learned frame selectors beat uniform sampling yet trail oracle evidence
frames.<br>
<strong>Evidence fusion:</strong> even with oracle frames, models often over-weight a single frame rather than
integrating dispersed cues, missing the right answer.
</p>
</div>
</section>
<!-- Task Families -->
<section class="results-section" id="tasks">
<h2>Task Families</h2>
<p class="section-description">
12 compositional tasks grouped into 4 reasoning families:<br>
1. Temporal Reasoning & Chronology (TR&C) <br>
2. Referring & Tracking (R&T)<br>
3. Global Consistency & Verification (GC&V)<br>
4. Multi-Entity Aggregation & Numeracy (MEA&N).
</p>
<div class="task-explorer">
<!-- Sidebar Menu -->
<div class="task-menu" id="task-menu">
<div class="task-family-group">
<h4>Temporal Reasoning & Chronology (TR&C)</h4>
<div class="task-menu-items">
<button class="task-menu-btn active" data-task="TSO">TSO: Temporal Shot Ordering</button>
<button class="task-menu-btn" data-task="MPDR">MPDR: Multi-Person Duration Reasoning</button>
<button class="task-menu-btn" data-task="ASII">ASII: Action Sequence Integrity & Identification</button>
</div>
</div>
<div class="task-family-group">
<h4>Referring & Tracking (R&T)</h4>
<div class="task-menu-items">
<button class="task-menu-btn" data-task="AGBI">AGBI: Appearance-Grounded Behavior Interactions</button>
<button class="task-menu-btn" data-task="AGAR">AGAR: Appearance-Grounded Attribute Recognition</button>
<button class="task-menu-btn" data-task="AGLT">AGLT: Appearance-Grounded Localization Trajectory</button>
</div>
</div>
<div class="task-family-group">
<h4>Global Consistency & Verification (GC&V)</h4>
<div class="task-menu-items">
<button class="task-menu-btn" data-task="FAM">FAM: False Action Memory</button>
<button class="task-menu-btn" data-task="SVA">SVA: Scene Verification Arrangement</button>
<button class="task-menu-btn" data-task="FOM">FOM: False Object Memory</button>
</div>
</div>
<div class="task-family-group">
<h4>Multi-Entity Aggregation & Numeracy (MEA&N)</h4>
<div class="task-menu-items">
<button class="task-menu-btn" data-task="MEGL">MEGL: Multi-Entities Grounding & Localization</button>
<button class="task-menu-btn" data-task="AC">AC: Action Counting</button>
<button class="task-menu-btn" data-task="RLPC">RLPC: Region-Localized People Counting</button>
</div>
</div>
</div>
<!-- Detail Viewer -->
<div class="task-viewer">
<div class="task-header">
<div class="task-title">
<span id="viewer-title">TSO: Temporal Shot Ordering</span>
<span class="task-badge" id="viewer-badge">Temporal Reasoning & Chronology (TR&C)</span>
</div>
<p class="task-desc" id="viewer-desc">
Arrange four shot descriptions into the correct chronological order using content cues alone.
</p>
<div class="task-abilities-box">
<span class="abilities-label">Abilities Tested:</span>
<span id="viewer-abilities">Understanding event order, high-level scene transitions, chronological
reconstruction using content cues</span>
</div>
</div>
<div class="task-content">
<img id="viewer-img" class="task-placeholder-img" src="" alt="Visual example">
<div class="placeholder-overlay" style="position: absolute; color: #64748b; font-weight: 500;">
<!-- Image Placeholder -->
</div>
</div>
</div>
</div>
</section>
<!-- MRFS -->
<section class="results-section" id="mrfs">
<h2>Evidential Requirement (MRFS)</h2>
<p class="section-description">
We quantify how much visual evidence a VideoQA item requires using the
<em>Minimum Required Frame-Set (MRFS)</em>: the <strong>smallest number of frames that must be integrated for a
model to answer correctly.</strong>
Higher MRFS indicates that questions are not solvable from a single salient snapshot and instead require
integrating temporally separated cues.
</p>
<div class="bar-chart-container" style="margin-top: 1rem;">
<!-- NExT-QA: 2.61 -->
<div class="bar-row">
<div class="bar-label">NExT-QA</div>
<div class="bar-track">
<div class="bar-fill" style="width: 43.5%;">2.61</div>
</div>
</div>
<!-- MVBench: 3.52 -->
<div class="bar-row">
<div class="bar-label">MVBench</div>
<div class="bar-track">
<div class="bar-fill" style="width: 58.7%;">3.52</div>
</div>
</div>
<!-- LongVideoBench: 4.07 -->
<div class="bar-row">
<div class="bar-label">LongVideoBench</div>
<div class="bar-track">
<div class="bar-fill" style="width: 67.8%;">4.07</div>
</div>
</div>
<!-- HERBench: 5.49 -->
<div class="bar-row">
<div class="bar-label"><strong>HERBench</strong></div>
<div class="bar-track">
<div class="bar-fill highlight" style="width: 91.5%;">5.49</div>
</div>
</div>
</div>
<p class="section-description" style="margin-top: 1.5rem;">
<strong>Cross-benchmark comparison.</strong>
HERBench exhibits the highest evidential requirement (mean MRFS = 5.49), exceeding LongVideoBench (4.07),
MVBench (3.52), and NExT-QA (2.61).
Notably, this higher MRFS is achieved despite a shorter average video duration than LongVideoBench, suggesting
the difficulty is driven by evidential density rather than video length alone.
This metric-level effect is consistent with the benchmark construction, which explicitly filters out
items that are solvable with too few frames.
</p>
<p class="section-description" style="font-size: 0.95em; opacity: 0.85; margin-top: 0.75rem;">
<strong>Protocol.</strong>
Because MRFS is defined with respect to a model <em>f</em>, a question-conditioned frame selector <em>r</em>,
and a frame budget <em>x</em>,
we report values under a standardized setup to enable apples-to-apples comparison across benchmarks.
</p>
</section>
<!-- Additional Figures -->
<section class="comparison-section" id="figures">
<h2>Benchmark statistics</h2>
<p class="bench-note">
The plots below give a quick snapshot of HERBench: where the questions come from, what they talk about, and how
they’re distributed across tasks.
</p>
<ul class="bench-bullets">
<li><span class="bench-label">Source mix</span> — question count by dataset.</li>
<li><span class="bench-label">Language</span> — frequent terms in the questions (word cloud).</li>
<li><span class="bench-label">Coverage</span> — question count across tasks.</li>
</ul>
<div class="figure-grid">
<div class="figure-card">
<div class="stats-layout">
<div class="stats-top-row">
<div class="stats-img-half">
<img src="assets/images/questions_per_dataset_pie.png" alt="Questions per Dataset" class="stats-img"
loading="lazy">
</div>
<div class="stats-img-half">
<img src="assets/images/wordcloud.jpg" alt="Wordcloud" class="stats-img" loading="lazy">
</div>
</div>
<div class="stats-img-full">
<img src="assets/images/task_bar_chart.png" alt="Task Distribution Bar Chart" class="stats-img"
loading="lazy">
</div>
</div>
</div>
</div>
</section>
<!-- Download & Evaluate -->
<section class="method-section" id="download">
<h2>Download and Evaluate</h2>
<div class="card-grid">
<!-- Card 1: Setup -->
<div class="card">
<h3>1. Setup</h3>
<p><strong>Prerequisites:</strong> Python 3.8-3.12, PyTorch 2.0+, and a GPU.</p>
<div class="code-block" style="margin-top: 1rem;">
<pre><code># Clone and install
git clone https://github.com/DanBenAmi/HERBench.git
cd HERBench
# Create env
conda env create -f environment.yml
conda activate herbench
pip install -e .</code></pre>
</div>
</div>
<!-- Card 2: Data -->
<div class="card">
<h3>2. Get Data</h3>
<p>Download video features and annotation files.</p>
<div class="code-block" style="margin-top: 1rem;">
<pre><code># Via Hugging Face (recommended)
python scripts/download_data.py --source huggingface
# Or via direct download
python scripts/download_data.py --source direct</code></pre>
</div>
</div>
<!-- Card 3: Evaluate -->
<div class="card" style="grid-column: 1 / -1;">
<h3>3. Evaluate</h3>
<p>Run inference and calculate metrics.</p>
<div class="code-grid" style="margin-top: 1rem;">
<div class="code-block">
<h4>Run Inference</h4>
<pre><code># Qwen2.5-VL (uniform frames)
python evaluation/run_evaluation.py \
model=qwen25vl frame_selector=uniform
# InternVL3.5 (BLIP frames)
python evaluation/run_evaluation.py \
model=internvl35 frame_selector=blip</code></pre>
</div>
<div class="code-block">
<h4>Calculate Metrics</h4>
<pre><code># Accuracy
python evaluation/calculate_accuracy.py \
--predictions results/predictions.json
# MRFS
python evaluation/calculate_mrfs.py \
model=qwen25vl frame_selector=blip</code></pre>
</div>
</div>
</div>
</div>
</section>
<!-- Citation -->
<section class="bibtex-section" id="citation">
<h2>Citation</h2>
<pre class="bibtex"><code>@misc{benami2025herbenchbenchmarkmultievidenceintegration,
title={HERBench: A Benchmark for Multi-Evidence Integration in Video Question Answering},
author={Dan Ben-Ami and Gabriele Serussi and Kobi Cohen and Chaim Baskin},
year={2025},
eprint={2512.14870},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2512.14870},
}</code></pre>
</section>
<!-- Footer -->
<footer class="page-footer">
<p>
© 2025 <a href="https://www.linkedin.com/in/gabriele-serussi/">Gabriele Serussi</a>,
<a href="https://www.linkedin.com/in/dan-ben-ami/">Dan Ben Ami</a>, and contributors. <br>
You are welcome to reuse or adapt this website's
<a href="https://github.com/GabrieleSerussi/HERBench">source code</a>
under the terms of the MIT License. <br>
We kindly ask that you include a link back to this page if you use it in your project.
</p>
</footer>
</div>
<script src="assets/js/main.js"></script>
</body>
</html>