-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
839 lines (794 loc) · 35.2 KB
/
index.html
File metadata and controls
839 lines (794 loc) · 35.2 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
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
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Look Twice (LoT) – Training-Free Evidence Highlighting in MLLMs</title>
<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=DM+Serif+Display:ital@0;1&family=DM+Sans:ital,opsz,wght@0,9..40,300;0,9..40,400;0,9..40,500;0,9..40,600;1,9..40,300;1,9..40,400&display=swap" rel="stylesheet" />
<style>
/* ── Reset & Base ────────────────────────────────── */
*, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
:root {
--bg: #ffffff;
--surface: #f8f8f6;
--border: #e8e6e0;
--text: #1a1a18;
--muted: #6b6b65;
--accent: #cc7aa2;
--accent2: #2c3e50;
--tag-bg: #f0ede8;
--tag-text: #3a3a35;
--radius: 10px;
--shadow: 0 2px 16px rgba(0,0,0,0.07);
--shadow-lg:0 8px 40px rgba(0,0,0,0.10);
--max-w: 960px;
--font-serif: 'DM Serif Display', Georgia, serif;
--font-sans: 'DM Sans', system-ui, sans-serif;
}
html { scroll-behavior: smooth; }
body {
font-family: var(--font-sans);
background: var(--bg);
color: var(--text);
font-size: 16px;
line-height: 1.7;
-webkit-font-smoothing: antialiased;
}
a { color: var(--accent); text-decoration: none; }
a:hover { text-decoration: underline; }
img { max-width: 100%; display: block; }
strong { font-weight: 600; }
/* ── Nav ─────────────────────────────────────────── */
nav {
position: sticky; top: 0; z-index: 100;
background: rgba(255,255,255,0.92);
backdrop-filter: blur(12px);
border-bottom: 1px solid var(--border);
padding: 0 2rem;
}
.nav-inner {
max-width: var(--max-w);
margin: 0 auto;
display: flex;
align-items: center;
justify-content: space-between;
height: 54px;
gap: 1.5rem;
}
.nav-logo {
font-family: var(--font-serif);
font-size: 1.2rem;
color: var(--text);
white-space: nowrap;
}
.nav-logo span { color: var(--accent); }
.nav-links {
display: flex;
gap: 1.4rem;
list-style: none;
flex-wrap: wrap;
}
.nav-links a {
color: var(--muted);
font-size: 0.85rem;
font-weight: 500;
letter-spacing: 0.02em;
text-transform: uppercase;
transition: color 0.2s;
}
.nav-links a:hover { color: var(--accent); text-decoration: none; }
/* ── Hero ────────────────────────────────────────── */
.hero {
padding: 5rem 2rem 3.5rem;
text-align: center;
max-width: var(--max-w);
margin: 0 auto;
}
.badge {
display: inline-flex;
align-items: center;
gap: 0.5rem;
background: var(--tag-bg);
color: var(--tag-text);
font-size: 0.78rem;
font-weight: 600;
letter-spacing: 0.06em;
text-transform: uppercase;
padding: 0.35rem 0.9rem;
border-radius: 50px;
margin-bottom: 1.6rem;
}
.badge::before {
content: '';
display: inline-block;
width: 7px; height: 7px;
background: var(--accent);
border-radius: 50%;
}
h1 {
font-family: var(--font-serif);
font-size: clamp(2rem, 4.5vw, 3rem);
line-height: 1.2;
color: var(--text);
margin-bottom: 1rem;
font-weight: 400;
}
h1 em { color: var(--accent); font-style: italic; }
.tagline {
font-size: 1.08rem;
color: var(--muted);
max-width: 640px;
margin: 0 auto 2.2rem;
line-height: 1.6;
}
.authors {
display: flex;
flex-wrap: wrap;
gap: 0.4rem 1.2rem;
justify-content: center;
margin-bottom: 0.6rem;
}
.author {
font-size: 0.97rem;
font-weight: 500;
color: var(--text);
}
.affil {
font-size: 0.85rem;
color: var(--muted);
margin-bottom: 2rem;
}
.btn-group {
display: flex;
gap: 0.7rem;
justify-content: center;
flex-wrap: wrap;
}
.btn {
display: inline-flex;
align-items: center;
gap: 0.4rem;
padding: 0.55rem 1.3rem;
border-radius: 50px;
font-size: 0.88rem;
font-weight: 600;
transition: all 0.2s;
border: 1.5px solid transparent;
}
.btn-primary {
background: var(--accent);
color: #fff;
border-color: var(--accent);
}
.btn-primary:hover { background: #e3b5c8; border-color: #e3b5c8; text-decoration: none; }
.btn-outline {
background: transparent;
color: var(--text);
border-color: var(--border);
}
.btn-outline:hover { border-color: var(--accent); color: var(--accent); text-decoration: none; }
/* ── Divider ─────────────────────────────────────── */
.divider {
max-width: var(--max-w);
margin: 0 auto;
height: 1px;
background: var(--border);
}
/* ── Sections ────────────────────────────────────── */
section {
padding: 4rem 2rem;
max-width: var(--max-w);
margin: 0 auto;
}
.section-label {
font-size: 0.72rem;
letter-spacing: 0.12em;
text-transform: uppercase;
font-weight: 600;
color: var(--accent);
margin-bottom: 0.5rem;
}
h2 {
font-family: var(--font-serif);
font-size: clamp(1.5rem, 3vw, 2rem);
font-weight: 400;
color: var(--text);
margin-bottom: 1.2rem;
line-height: 1.25;
}
.prose {
color: #3a3a38;
max-width: 720px;
font-size: 1rem;
line-height: 1.75;
}
.prose p + p { margin-top: 1em; }
/* ── Teaser image ────────────────────────────────── */
.teaser {
margin: 2.5rem auto 0;
border-radius: var(--radius);
overflow: hidden;
border: 1px solid var(--border);
box-shadow: var(--shadow-lg);
background: var(--surface);
}
.img-placeholder {
width: 100%;
background: var(--surface);
border: 2px dashed var(--border);
border-radius: var(--radius);
display: flex;
flex-direction: column;
align-items: center;
justify-content: center;
gap: 0.5rem;
color: var(--muted);
font-size: 0.85rem;
font-weight: 500;
min-height: 280px;
}
.img-placeholder svg { opacity: 0.35; }
.img-caption {
text-align: center;
font-size: 0.85rem;
color: var(--muted);
margin-top: 0.9rem;
font-style: italic;
}
/* ── Stats bar ───────────────────────────────────── */
.stats {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(140px, 1fr));
gap: 1px;
background: var(--border);
border: 1px solid var(--border);
border-radius: var(--radius);
overflow: hidden;
margin-top: 3rem;
}
.stat {
background: var(--bg);
padding: 1.5rem 1.2rem;
text-align: center;
}
.stat-val {
font-family: var(--font-serif);
font-size: 2rem;
color: var(--accent);
line-height: 1;
margin-bottom: 0.3rem;
}
.stat-label {
font-size: 0.78rem;
color: var(--muted);
line-height: 1.3;
}
/* ── Two-pass pipeline ───────────────────────────── */
.pipeline {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 1.5rem;
margin-top: 2.5rem;
}
@media (max-width: 640px) { .pipeline { grid-template-columns: 1fr; } }
.pipe-step {
background: var(--surface);
border: 1px solid var(--border);
border-radius: var(--radius);
padding: 1.6rem;
position: relative;
}
.pipe-num {
font-family: var(--font-serif);
font-size: 3rem;
color: var(--border);
position: absolute;
top: 0.5rem; right: 1rem;
line-height: 1;
user-select: none;
}
.pipe-step h3 {
font-size: 0.92rem;
font-weight: 600;
letter-spacing: 0.03em;
margin-bottom: 0.5rem;
color: var(--text);
}
.pipe-step p {
font-size: 0.88rem;
color: var(--muted);
line-height: 1.6;
}
.pipe-tag {
display: inline-block;
background: var(--tag-bg);
color: var(--accent);
font-size: 0.7rem;
font-weight: 700;
letter-spacing: 0.06em;
text-transform: uppercase;
padding: 0.15rem 0.55rem;
border-radius: 4px;
margin-bottom: 0.6rem;
}
/* ── Method figure full ──────────────────────────── */
.full-fig {
margin: 2.5rem 0 0;
border: 1px solid var(--border);
border-radius: var(--radius);
overflow: hidden;
box-shadow: var(--shadow);
background: var(--surface);
}
/* ── Component cards (ablation highlights) ───────── */
.components {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 1rem;
margin-top: 2rem;
}
@media (max-width: 560px) { .components { grid-template-columns: 1fr; } }
.comp-card {
border: 1px solid var(--border);
border-radius: var(--radius);
padding: 1.4rem 1.4rem 1.2rem;
background: var(--bg);
transition: box-shadow 0.2s, border-color 0.2s;
}
.comp-card:hover { box-shadow: var(--shadow); border-color: #ccc; }
.comp-icon {
font-size: 1.4rem;
margin-bottom: 0.6rem;
}
.comp-card h3 {
font-size: 0.93rem;
font-weight: 600;
margin-bottom: 0.4rem;
}
.comp-card p {
font-size: 0.85rem;
color: var(--muted);
line-height: 1.55;
}
/* ── Results table ───────────────────────────────── */
.table-wrap {
overflow-x: auto;
margin-top: 2rem;
border: 1px solid var(--border);
border-radius: var(--radius);
box-shadow: var(--shadow);
}
table {
width: 100%;
border-collapse: collapse;
font-size: 0.84rem;
}
thead th {
background: var(--surface);
padding: 0.75rem 1rem;
font-size: 0.72rem;
letter-spacing: 0.05em;
text-transform: uppercase;
color: var(--muted);
font-weight: 600;
text-align: left;
border-bottom: 1px solid var(--border);
white-space: nowrap;
}
tbody td {
padding: 0.65rem 1rem;
border-bottom: 1px solid var(--border);
color: var(--text);
}
tbody tr:last-child td { border-bottom: none; }
tbody tr:hover td { background: var(--surface); }
.row-lot td { background: #fef9f8; }
.row-lot td:first-child { border-left: 3px solid var(--accent); padding-left: calc(1rem - 3px); }
.delta {
display: inline-block;
background: #edfaf2;
color: #1a7a3a;
font-size: 0.75rem;
font-weight: 700;
padding: 0.1rem 0.45rem;
border-radius: 4px;
}
/* ── Qualitative grid ────────────────────────────── */
.qual-grid {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 1.2rem;
margin-top: 2rem;
}
@media (max-width: 600px) { .qual-grid { grid-template-columns: 1fr; } }
.qual-card {
border: 1px solid var(--border);
border-radius: var(--radius);
overflow: hidden;
background: var(--surface);
}
.qual-card .img-placeholder { min-height: 200px; border-radius: 0; border: none; border-bottom: 1px solid var(--border); }
.qual-card-body { padding: 0.9rem 1rem; }
.qual-card-body p { font-size: 0.82rem; color: var(--muted); }
/* ── BibTeX ──────────────────────────────────────── */
.bibtex-block {
background: var(--surface);
border: 1px solid var(--border);
border-radius: var(--radius);
padding: 1.6rem 1.8rem;
margin-top: 1.8rem;
position: relative;
}
pre {
font-family: 'JetBrains Mono', 'Fira Code', monospace;
font-size: 0.8rem;
line-height: 1.8;
color: var(--text);
white-space: pre-wrap;
word-break: break-all;
}
.copy-btn {
position: absolute;
top: 1rem; right: 1rem;
background: var(--bg);
border: 1px solid var(--border);
border-radius: 6px;
padding: 0.3rem 0.75rem;
font-size: 0.75rem;
font-family: var(--font-sans);
font-weight: 600;
color: var(--muted);
cursor: pointer;
transition: all 0.2s;
}
.copy-btn:hover { border-color: var(--accent); color: var(--accent); }
/* ── Footer ──────────────────────────────────────── */
footer {
border-top: 1px solid var(--border);
padding: 2.5rem 2rem;
text-align: center;
font-size: 0.82rem;
color: var(--muted);
line-height: 1.8;
}
footer a { color: var(--muted); }
footer a:hover { color: var(--accent); }
/* ── Scroll-in animations ────────────────────────── */
.reveal {
opacity: 0;
transform: translateY(22px);
transition: opacity 0.55s ease, transform 0.55s ease;
}
.reveal.visible { opacity: 1; transform: none; }
/* ── Responsive ──────────────────────────────────── */
@media (max-width: 680px) {
.nav-links { display: none; }
h1 { font-size: 1.75rem; }
}
</style>
</head>
<body>
<!-- ── Navigation ──────────────────────────────────────── -->
<nav>
<div class="nav-inner">
<span class="nav-logo">Look <span>Twice</span></span>
<ul class="nav-links">
<li><a href="#abstract">Abstract</a></li>
<li><a href="#method">Method</a></li>
<li><a href="#results">Results</a></li>
<li><a href="#qualitative">Qualitative</a></li>
<li><a href="#bibtex">BibTeX</a></li>
</ul>
</div>
</nav>
<!-- ── Hero ────────────────────────────────────────────── -->
<header class="hero">
<h1>Look Twice: Training-Free<br><em>Evidence Highlighting</em><br>in Multimodal Large Language Models</h1>
<p class="tagline">A lightweight inference-time framework that improves how pretrained MLLMs utilize multimodal evidence — no fine-tuning, no architectural changes.</p>
<div class="authors">
<span class="author">Marco Morini</span>
<span class="author">Sara Sarto</span>
<span class="author">Marcella Cornia</span>
<span class="author">Lorenzo Baraldi</span>
</div>
<p class="affil">University of Modena and Reggio Emilia</p>
<div class="btn-group">
<a href="https://arxiv.org/abs/2604.01280" class="btn btn-primary">
<!-- arxiv icon -->
<svg width="16" height="16" viewBox="0 0 24 24" fill="currentColor"><path d="M12 2C6.477 2 2 6.477 2 12s4.477 10 10 10 10-4.477 10-10S17.523 2 12 2zm-1 14H9V8h2v8zm4 0h-2V8h2v8z"/></svg>
arXiv Paper
</a>
<a href="https://github.com/aimagelab/LoT" class="btn btn-outline">
<svg width="16" height="16" viewBox="0 0 24 24" fill="currentColor"><path d="M12 2A10 10 0 0 0 2 12c0 4.42 2.87 8.17 6.84 9.5.5.08.66-.23.66-.5v-1.69c-2.77.6-3.36-1.34-3.36-1.34-.46-1.16-1.11-1.47-1.11-1.47-.91-.62.07-.6.07-.6 1 .07 1.53 1.03 1.53 1.03.87 1.52 2.34 1.07 2.91.83.09-.65.35-1.09.63-1.34-2.22-.25-4.55-1.11-4.55-4.92 0-1.11.38-2 1.03-2.71-.1-.25-.45-1.29.1-2.64 0 0 .84-.27 2.75 1.02.79-.22 1.65-.33 2.5-.33.85 0 1.71.11 2.5.33 1.91-1.29 2.75-1.02 2.75-1.02.55 1.35.2 2.39.1 2.64.65.71 1.03 1.6 1.03 2.71 0 3.82-2.34 4.66-4.57 4.91.36.31.69.92.69 1.85V21c0 .27.16.59.67.5C19.14 20.16 22 16.42 22 12A10 10 0 0 0 12 2z"/></svg>
GitHub Code
</a>
<a href="https://aimagelab.ing.unimore.it/" class="btn btn-outline">AImageLab</a>
</div>
</header>
<div class="divider"></div>
<!-- ── Abstract ─────────────────────────────────────────── -->
<section id="abstract">
<div class="section-label">Abstract</div>
<h2>From noisy evidence to precise answers</h2>
<div class="prose">
<p>Answering questions about images often requires combining visual understanding with external knowledge. Multimodal Large Language Models (MLLMs) provide a natural framework for this setting, but they often <strong>struggle to identify the most relevant visual and textual evidence</strong> when answering knowledge-intensive queries.</p>
<p>In such scenarios, models must integrate visual cues with retrieved textual evidence that is often noisy or only partially relevant, while also localizing fine-grained visual information in the image. We introduce <strong>Look Twice (LoT)</strong>, a training-free inference-time framework that improves how pretrained MLLMs utilize multimodal evidence.</p>
<p>We exploit the model attention patterns to estimate which visual regions and retrieved textual elements are relevant to a query, then generate the answer conditioned on this highlighted evidence. The selected cues are highlighted through <strong>lightweight prompt-level markers</strong> that encourage the model to re-attend to the relevant evidence during generation — with negligible computational overhead.</p>
</div>
<!-- Teaser placeholder -->
<div class="teaser reveal">
<div class="img-placeholder" style="min-height:360px;">
<!-- <svg width="48" height="48" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.2"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span>Placeholder — insert teaser figure here (e.g., Fig. 1 from the paper)</span>
<span style="font-size:0.75rem; opacity:0.7;">Recommended: images/teaser.png · ~1200×500 px</span> -->
<img src="images/03_qualitatives-1.png" alt="Teaser figure illustrating the Look Twice (LoT) framework" />
</div>
</div>
<p class="img-caption">Figure 1. Look Twice (LoT): a two-pass inference strategy that highlights query-relevant visual regions and textual evidence to improve multimodal reasoning.</p>
<!-- Stats bar -->
<div class="stats reveal">
<div class="stat">
<div class="stat-val">+5.3</div>
<div class="stat-label">Max avg. gain on KB-VQA benchmarks</div>
</div>
<div class="stat">
<div class="stat-val">0</div>
<div class="stat-label">Additional training or fine-tuning</div>
</div>
<div class="stat">
<div class="stat-val">+1</div>
<div class="stat-label">Single extra token in first pass</div>
</div>
<div class="stat">
<div class="stat-val">10</div>
<div class="stat-label">Models evaluated across 3 scales</div>
</div>
</div>
</section>
<div class="divider"></div>
<!-- ── Method ───────────────────────────────────────────── -->
<section id="method">
<div class="section-label">Method</div>
<h2>A two-pass inference strategy</h2>
<div class="prose">
<p>LoT operates entirely at inference time. The model looks at the input twice: a first lightweight generation step produces a single token, used to inspect internal attention patterns. A second generation pass produces the final answer conditioned on highlighted evidence.</p>
</div>
<div class="pipeline reveal">
<div class="pipe-step">
<span class="pipe-num">01</span>
<div class="pipe-tag">First pass</div>
<h3>Attention Analysis</h3>
<p>Generate a single token to read the model's internal attention patterns. Extract object-to-visual attention to locate relevant image regions, and last-to-context attention to score retrieved sentences.</p>
</div>
<div class="pipe-step">
<span class="pipe-num">02</span>
<div class="pipe-tag">Filtering</div>
<h3>Attention Sink Filtering</h3>
<p>Identify and suppress spurious visual tokens acting as attention sinks — tokens that attract disproportionate attention regardless of semantic relevance — using hidden-state activation statistics.</p>
</div>
<div class="pipe-step">
<span class="pipe-num">03</span>
<div class="pipe-tag">Localization</div>
<h3>Visual Evidence Selection</h3>
<p>Aggregate filtered attention into a 2D map and extract a bounding box via weighted centroid and spread. The cropped region is inserted into the prompt with special markers.</p>
</div>
<div class="pipe-step">
<span class="pipe-num">04</span>
<div class="pipe-tag">Second pass</div>
<h3>Highlighted Generation</h3>
<p>The final answer is generated with the original image replaced by the cropped evidence and the most relevant retrieved sentence wrapped in importance markers, guiding the model's attention.</p>
</div>
</div>
<!-- Method figure placeholder -->
<div class="full-fig reveal">
<div class="img-placeholder" style="min-height:320px; border-radius:0; border:none;">
<!-- <svg width="48" height="48" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.2"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span>Placeholder — insert method overview figure (Fig. 1 from paper)</span>
<span style="font-size:0.75rem; opacity:0.7;">Recommended: images/method.png · ~1200×600 px</span> -->
<img src="images/00_first-1.png" alt="Overview of the LoT pipeline" />
</div>
</div>
<p class="img-caption">Figure 2. Overview of the LoT pipeline, showing textual evidence selection (attention matrix → sentence highlighting) and visual evidence selection (sink filtering → bounding box extraction).</p>
<!-- Components -->
<div style="margin-top:3rem;">
<div class="section-label">Components</div>
<h2>What makes LoT work</h2>
</div>
<div class="components reveal">
<div class="comp-card">
<div class="comp-icon">🔍</div>
<h3>Self-Guided Visual Evidence</h3>
<p>Object-conditioned attention between question tokens and visual tokens produces a query-specific relevance map, explicitly capturing how the queried object interacts with the visual input.</p>
</div>
<div class="comp-card">
<div class="comp-icon">🚫</div>
<h3>Attention Sink Filtering</h3>
<p>Tokens with disproportionately high hidden-state activations in sink dimensions (identified from the base LLM) are suppressed, yielding cleaner grounding maps without modifying the model.</p>
</div>
<div class="comp-card">
<div class="comp-icon">📦</div>
<h3>Weighted Centroid Bounding Box</h3>
<p>The attention map is converted into a precise spatial region by computing the attention-weighted centroid and standard deviation, outperforming min-max and morphological alternatives.</p>
</div>
<div class="comp-card">
<div class="comp-icon">📝</div>
<h3>Self-Guided Textual Evidence</h3>
<p>Last-token-to-context attention aggregated across deep decoder layers identifies the single most relevant retrieved sentence, which is wrapped with importance markers during generation.</p>
</div>
</div>
<!-- Attention map filtering figure placeholder -->
<div style="margin-top:2.5rem;" class="reveal">
<div class="img-placeholder" style="min-height:260px;">
<!-- <svg width="48" height="48" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.2"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span>Placeholder — attention map filtering figure (Fig. 2 from paper)</span>
<span style="font-size:0.75rem; opacity:0.7;">Recommended: images/filtering.png · ~1100×380 px</span> -->
<img src="images/01_fig-1.png" alt="Qualitative examples of attention sink filtering, showing raw and filtered attention maps" />
</div>
<p class="img-caption">Figure 3. Qualitative examples of attention sink filtering. Raw maps (center) contain scattered activations; filtered maps (right) are tightly concentrated around the target object.</p>
</div>
</section>
<div class="divider"></div>
<!-- ── Results ───────────────────────────────────────────── -->
<section id="results">
<div class="section-label">Results</div>
<h2>Consistent gains across models & benchmarks</h2>
<div class="prose">
<p>LoT is evaluated on four KB-VQA benchmarks (E-VQA, InfoSeek, OVEN, ViQuAE) across ten off-the-shelf MLLMs in a zero-shot setting. Gains range from <strong>+1.1 to +5.3 average accuracy</strong> depending on the backbone, with no additional training.</p>
</div>
<!-- Results table (excerpt – key models) -->
<div class="table-wrap reveal">
<table>
<thead>
<tr>
<th>Model</th>
<th>E-VQA All</th>
<th>InfoSeek All</th>
<th>OVEN All</th>
<th>ViQuAE</th>
<th>Avg</th>
</tr>
</thead>
<tbody>
<tr><td colspan="6" style="font-size:0.72rem;letter-spacing:0.05em;text-transform:uppercase;color:var(--muted);padding:0.5rem 1rem;background:var(--surface);">Small Models</td></tr>
<tr><td>Qwen2.5-VL-3B</td><td>27.8</td><td>22.4</td><td>11.5</td><td>22.9</td><td>21.2</td></tr>
<tr class="row-lot"><td><strong>+ LoT (Ours)</strong></td><td>30.4</td><td>25.2</td><td>18.3</td><td>27.8</td><td>25.5 <span class="delta">+4.3</span></td></tr>
<tr><td>InternVL3.5-4B</td><td>26.2</td><td>28.9</td><td>10.8</td><td>36.4</td><td>25.6</td></tr>
<tr class="row-lot"><td><strong>+ LoT (Ours)</strong></td><td>28.7</td><td>33.2</td><td>11.5</td><td>45.6</td><td>29.8 <span class="delta">+4.2</span></td></tr>
<tr><td colspan="6" style="font-size:0.72rem;letter-spacing:0.05em;text-transform:uppercase;color:var(--muted);padding:0.5rem 1rem;background:var(--surface);">Medium Models</td></tr>
<tr><td>Qwen2-VL-7B</td><td>22.9</td><td>24.4</td><td>11.1</td><td>33.0</td><td>22.9</td></tr>
<tr class="row-lot"><td><strong>+ LoT (Ours)</strong></td><td>25.6</td><td>29.9</td><td>16.6</td><td>40.8</td><td>28.2 <span class="delta">+5.3</span></td></tr>
<tr><td>Qwen3-VL-8B</td><td>35.0</td><td>29.7</td><td>17.7</td><td>43.7</td><td>31.5</td></tr>
<tr class="row-lot"><td><strong>+ LoT (Ours)</strong></td><td>36.4</td><td>32.8</td><td>19.6</td><td>51.0</td><td>35.0 <span class="delta">+3.5</span></td></tr>
<tr><td colspan="6" style="font-size:0.72rem;letter-spacing:0.05em;text-transform:uppercase;color:var(--muted);padding:0.5rem 1rem;background:var(--surface);">Large Models</td></tr>
<tr><td>InternVL3.5-38B</td><td>31.6</td><td>33.1</td><td>20.2</td><td>51.5</td><td>34.1</td></tr>
<tr class="row-lot"><td><strong>+ LoT (Ours)</strong></td><td>33.5</td><td>36.8</td><td>18.5</td><td>61.0</td><td>37.5 <span class="delta">+3.1</span></td></tr>
</tbody>
</table>
</div>
<p class="img-caption">Table 1. Performance on KB-VQA benchmarks. LoT consistently improves all backbones without any training.</p>
<!-- Results chart placeholder -->
<div style="margin-top:2rem;" class="reveal">
<div class="img-placeholder" style="min-height:260px;">
<!-- <svg width="48" height="48" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.2"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span>Placeholder — bar chart / ablation figure (Fig. 4 from paper)</span>
<span style="font-size:0.75rem; opacity:0.7;">Recommended: images/ablation.png · ~900×340 px</span> -->
<img src="images/04_abl_combined-1.png" alt="Bar charts showing LoT performance as the number of retrieved passages varies and with oracle evidence" />
</div>
<p class="img-caption">Figure 4. Performance on E-VQA as the number of retrieved passages varies (left) and with oracle evidence (right). LoT consistently outperforms zero-shot baselines in all settings.</p>
</div>
<!-- Generalization section -->
<div style="margin-top:3.5rem;">
<div class="section-label">Generalization</div>
<h2>Visual highlighting alone helps general MLLMs</h2>
<div class="prose">
<p>When applied with only visual cues (no retrieved text), LoT still improves performance on <strong>vision-centric</strong> (RealWorldQA, V-Star), <strong>OCR</strong> (TextVQA, OCRBench, ChartQA), and <strong>hallucination</strong> benchmarks (POPE, AMBER-D), demonstrating that improved visual grounding benefits multimodal reasoning broadly.</p>
</div>
</div>
</section>
<!-- <div class="divider"></div> -->
<!-- ── Qualitative ───────────────────────────────────────── -->
<section id="qualitative">
<!-- <div class="section-label">Qualitative Results</div>
<h2>Seeing what matters</h2>
<div class="prose">
<p>LoT reliably identifies the precise image region containing the object referenced in the question and highlights the retrieved sentence that directly provides the answer.</p>
</div>
<div class="qual-grid reveal">
<div class="qual-card">
<div class="img-placeholder">
<svg width="32" height="32" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.3"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span style="font-size:0.75rem;">images/qual_1.png</span>
</div>
<div class="qual-card-body">
<p><strong>Q:</strong> What color are the undersides of this female butterfly's wings?<br><strong>LoT:</strong> Yellowish with black speckles ✓</p>
</div>
</div>
<div class="qual-card">
<div class="img-placeholder">
<svg width="32" height="32" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.3"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span style="font-size:0.75rem;">images/qual_2.png</span>
</div>
<div class="qual-card-body">
<p><strong>Q:</strong> Which animals eat the fruits / leaves of this plant?<br><strong>LoT:</strong> Bees, hummingbirds, and lepidoptera ✓</p>
</div>
</div>
<div class="qual-card">
<div class="img-placeholder">
<svg width="32" height="32" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.3"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span style="font-size:0.75rem;">images/qual_3.png</span>
</div>
<div class="qual-card-body">
<p><strong>Q:</strong> Where is this insect native?<br><strong>LoT:</strong> Brazil to central Mexico ✓</p>
</div>
</div>
<div class="qual-card">
<div class="img-placeholder">
<svg width="32" height="32" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.3"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span style="font-size:0.75rem;">images/qual_4.png</span>
</div>
<div class="qual-card-body">
<p><strong>Q:</strong> After what or who was this species named?<br><strong>LoT:</strong> John Edwards Holbrook ✓</p>
</div>
</div>
</div>
<p class="img-caption">Figure 5. LoT correctly localizes the relevant image region (red bounding box) and highlights the evidence sentence (yellow) that contains the answer.</p> -->
<!-- Hallucination figure placeholder -->
<div style="margin-top:-6rem;" class="reveal">
<div class="img-placeholder" style="min-height:220px;">
<!-- <svg width="48" height="48" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.2"><rect x="3" y="3" width="18" height="18" rx="2"/><circle cx="8.5" cy="8.5" r="1.5"/><path d="m21 15-5-5L5 21"/></svg>
<span>Placeholder — hallucination benchmark examples (Fig. 7 from paper)</span>
<span style="font-size:0.75rem; opacity:0.7;">Recommended: images/hallucination.png · ~1100×400 px</span> -->
<img src="images/06_hall_qualitatives_cropped (1)-1.png" alt="Qualitative examples of LoT on hallucination benchmarks, showing improved grounding and correct answers" />
</div>
<p class="img-caption">Figure 6. LoT on POPE and AMBER: visual grounding guides the model toward the correct yes/no answer about object presence.</p>
</div>
</section>
<div class="divider"></div>
<!-- ── BibTeX ────────────────────────────────────────────── -->
<section id="bibtex">
<div class="section-label">Citation</div>
<h2>BibTeX</h2>
<div class="bibtex-block">
<button class="copy-btn" onclick="copyBib()">Copy</button>
<pre id="bibtex-text">@article{morini2026looktwice,
title = {{Look Twice: Training-Free Evidence Highlighting in Multimodal Large Language Models}},
author = {Morini, Marco and Sarto, Sara and Cornia, Marcella and Baraldi, Lorenzo},
journal = {arXiv preprint arXiv:2604.01280},
year = {2026}
}</pre>
</div>
</section>
<div class="divider"></div>
<!-- ── Footer ────────────────────────────────────────────── -->
<footer>
<p>
<a href="https://aimagelab.ing.unimore.it/">AImageLab</a> ·
University of Modena and Reggio Emilia
</p>
<p style="margin-top:0.4rem;">
This work was supported by the EU Horizon project ELLIOT (No. 101214398)
and the EuroHPC JU project MINERVA (GA No. 101182737).
</p>
</footer>
<!-- ── Scripts ───────────────────────────────────────────── -->
<script>
// Scroll reveal
const revealEls = document.querySelectorAll('.reveal');
const obs = new IntersectionObserver(entries => {
entries.forEach(e => { if (e.isIntersecting) { e.target.classList.add('visible'); obs.unobserve(e.target); } });
}, { threshold: 0.08 });
revealEls.forEach(el => obs.observe(el));
// Copy BibTeX
function copyBib() {
const text = document.getElementById('bibtex-text').textContent;
navigator.clipboard.writeText(text).then(() => {
const btn = document.querySelector('.copy-btn');
btn.textContent = 'Copied!';
setTimeout(() => btn.textContent = 'Copy', 2000);
});
}
</script>
</body>
</html>