Advertisers routinely create multiple duration variants of ads to meet marketing budgets and viewer preferences, a labor-intensive and costly process. While multimodal large language models (MLLMs) excel at general video understanding, their ability to perform specialized editorial tasks remains unexplored. We introduce AdShot, a benchmark for evaluating MLLMs on shot selection for ad clipping. It contains 823 professionally edited ad pairs (30-sec sources and 15-sec edits) across 194 brands and 17 industries, annotated with shot-level boundaries and content features (action, emotion, information, imagery). We evaluate 13 state-of-the-art video-only MLLMs and 2 audio-visual MLLMs (7B-72B parameters) on shot selection accuracy (Precision, Recall, F1), duration constraints (15s), and content preservation, including an audio ablation. We find: (1) model architecture matters more than scale, with smaller models often matching or exceeding larger ones; performance is strongest for target ads with 6-10 shots; (2) models exhibit a mean absolute duration error of 5.39s, with most tending to overshoot the 15s target, though they can still substantially reduce human editing effort; (3) all models exhibit systematic content biases, favoring information-rich and emotion-focused shots while under-selecting action and imagery. We publicly release our benchmark to facilitate research at the intersection of multimodal AI and computational advertising.
Figures & tables
Figure 1: Example of the shot selection task: From a 31-sec Subaru ad from YouTube (shots indexed in red on the right panel; we show the first frame of each shot for better representation), professional editors select and order shots (blue indices) to create a 16-sec version promoting the same product (CrossTrek).
Benchmark
Domain
Task
Videos
Length
Key Distinction
VideoAds Zhang et al. (2025)
Ads
Open-ended QA
3,153
15–60s
QA on ad content
AdsQA Long et al. (2025)
Ads
Open-ended QA
600
5–90s
Reasons about implicit ad meanings
AdSum Xie et al. (2026)
Ads
Shot Selection
204
15s, 30s
Supervised audio-visual model; no MLLM evaluation
Video-MME Fu et al. (2025)
General
Multiple-choice QA
900
11s–1h
Broad multimodal understanding
MVBench Li et al. (2023)
General
Open-ended QA
4,000
10s–10min
Temporal reasoning across events
MLVU Zhou et al. (2025)
General
Multiple-choice QA
1,500+
Hours
Long-form video reasoning
Table 1: Comparison between AdShot and existing video understanding benchmarks.
Figure 2: Methodological overview of long-short ad pair construction. Given videos from a brand’s YouTube channel, we first apply AutoShot ( Zhu et al., 2023 ) for shot boundary detection, then extract shot embeddings using Swin3D ( Liu et al., 2022 ) . We cluster the embeddings using DBSCAN to match each shot in the 15-sec ad to its corresponding shot in the 30-sec ad. Throughout the figure, blue (red) elements (e.g., digits and dots) represent the 15-sec (30-sec) ad.
Figure 4
Figure 4: MLLM evaluation pipeline for ad shot selection. Keyframes from shots of a 30-sec ad, along with metadata (e.g., durations) and task instructions specifying the clipping objective, are fed to the model. The model outputs selected shots with reasoning, which are evaluated for accuracy and content preservation.
Model
Family
Params
Modality
LLaVA-OneVision ( Li et al., 2024a )
LLaVA
7B, 72B
V
Qwen2.5-VL ( Bai et al., 2025b )
Qwen
7B, 32B
V
Qwen3-VL ( Bai et al., 2025a )
Qwen
8B, 32B
V
InternVL2.5 ( Chen et al., 2024 )
InternVL
8B, 26B
V
InternVL3.5 ( Zhu et al., 2025 )
InternVL
8B, 38B
V
Aria ( Li et al., 2024b )
MoE
25.3B
V
Table 3: Evaluated models, parameter scales, and input modalities. V = video-only; A+V = audio + video.
Model
Precision ↑
Recall ↑
F1 ↑
MDE ↓
MDE pos↓
MDE neg↓
Random baseline
0.569
0.541
0.503
8.18
9.01
7.16
Video-only inputs
Qwen2.5-VL 7B
0.613
0.649
0.582
7.82
5.92
1.90
LLaVA-OneVision 7B
0.578
0.601
0.539
8.68
6.41
2.27
Qwen3-VL 8B
0.618
0.580
0.576
4.08
2.99
1.10
InternVL2.5 8B
0.594
0.698
0.613
8.30
7.98
0.32
Table 4: Performance on shot selection (1 FPS). Random baseline samples a random subset size N , then picks N shots uniformly at random (Appendix D ). Bold marks the best per column across all evaluated models. ↑ / ↓ indicate higher/lower is better.
Model
F1 ↑
MDE (s) ↓
SC < 6
6 ≤ SC ≤ 10
SC > 10
SC < 6
6 ≤ SC ≤ 10
SC > 10
Video-only inputs
Qwen2.5-VL 7B
0.609
0.706
0.719
9.95
12.44
14.05
LLaVA-OneVision 7B
0.534
0.467
0.381
4.46
3.40
5.21
Qwen3-VL 8B
0.595
0.614
0.640
7.39
8.11
10.40
InternVL2.5 8B
0.480
0.450
0.360
3.32
4.06
6.45
Table 5: Shot selection performance by targeted shot count (SC) at fps=1. F1 and MDE for 15-sec ads with < 6, 6–10, and > 10 shots. Shaded cells highlight per-model best F1 and lowest MDE across bins; bold marks the best per column.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Category
Question
Action
Q1: Is there a call to find out more information about the product or the brand (e.g., shop online, visit the Web)?
Q2: Is there online contact information provided (e.g., URL, website)?
Q3: Is there a visual or verbal call to purchase (e.g., buy now, order now)?
Q4: Does the ad portray a sense of urgency to act (e.g., buy before sales ends, order before ends)?
Q5: Is there an incentive to buy (e.g., a discount, a coupon, a sale or "limited time offer")?
Q6: Is there offline contact information provided (e.g., phone, mail, store location)?
Appendix
Table 6: Annotation Questions of Ad Content Feature by Category
Figure 5: Content Feature Divergence. MLLMs over-select emotion- and information-focused shots while under-selecting action- and imagery-driven content, indicating systematic biases relative to professional edits.
Figure 6: Ablation on temporal sampling. Using 3 or 5 frames per shot instead of 1 FPS does not consistently improve MLLM performance, indicating limited benefit from denser temporal input.
Model
fps=1
fps=3
fps=5
F1 ↑
MDE ↓
F1 ↑
MDE ↓
F1 ↑
MDE ↓
Video-only inputs
Qwen2.5-VL 7B
0.582
7.82
0.576
9.79
0.643
10.91
LLaVA-OneVision 7B
0.539
8.68
0.540
8.80
0.541
8.41
Qwen3-VL 8B
0.576
4.08
0.573
4.46
0.586
4.96
InternVL2.5 8B
0.613
8.30
0.390
11.09
—
—
Appendix
Table 7: Ablation on FPS. Shot selection performance under different temporal sampling densities. ↑ / ↓ indicate higher/lower is better. Shaded cells highlight, for each model, the best performance across the three FPS values; bold marks the best across all models. Some models (e.g., InternVL2.5) are excluded from 5 FPS due to token-length constraints.
Industry
Prec.
Rec.
F1
Food & Beverages
0.707
0.674
0.674
Cosmetic
0.618
0.699
0.634
Restaurants
0.664
0.631
0.634
Services
0.619
0.657
0.625
Health
0.638
0.641
0.622
Telecommunications
0.615
0.661
0.621
Appendix
Table 8: Per-industry shot-selection performance of Qwen3-VL 32B (1 FPS), sorted by F1. Shaded cells are above the overall average (Precision 0.636, Recall 0.618, F1 0.611). Industry labels follow Table 2 .
Model
Input
Prec.
Rec.
F1
MDE
MDE neg
MDE pos
MiniCPM-o 2.6 8B
Audio+Video
0.581
0.841
0.669
10.06
1.72
10.50
Video-only
0.588
0.778
0.644
8.38
2.28
9.60
Δ (audio)
−0.007
+0.063
+0.025
+1.68
−0.56
+0.90
Qwen2.5-Omni 7B
Audio+Video
0.615
0.838
0.668
11.18
3.16
12.68
Video-only
0.595
0.660
0.585
7.75
3.47
9.81
Δ (audio)
+0.020
+0.178
+0.083
+3.43
−0.31
+2.87
Appendix
Table 9: Audio ablation: same model, same 823 AdShot pairs, with and without the audio channel. Audio+Video rows are from Table 4 ; Video-only rows are this ablation. Δ is the audio contribution (Audio+Video − Video-only). MDE neg / MDE pos are the mean under-/over-shoot in seconds.
Industry
F1 +A
F1 −A
Δ F1
Rec +A
Rec −A
Δ Rec
Retail
0.659
0.586
+0.073
0.825
0.639
+0.186
Legal
0.644
0.602
+0.042
0.822
0.689
+0.133
Services
0.648
0.567
+0.082
0.883
0.716
+0.167
Government
0.669
0.563
+0.106
0.825
0.611
+0.213
Health
0.640
0.575
+0.064
0.817
0.637
+0.180
Automotive
0.727
0.571
+0.156
0.920
0.604
+0.316
Appendix
Table 10: Per-industry audio ablation for Qwen2.5-Omni 7B across all 17 industries, sorted by volume. “ + A” is audio+video, “ − A” is video-only; Δ is the audio contribution. Shaded Δ F1 cells are positive: audio improves F1 in 14 of 17 industries and recall in 16 of 17.
Figure 7: Example 1 — Camping World (Automotive / Vehicles), 22 shots. Blue: Qwen3-VL 32B selected; yellow: golden shot (ground truth).
Figure 8: Example 2 — Cologuard (Health), 16 shots. Blue: Qwen3-VL 32B selected; yellow: golden shot (ground truth).
Figure 9: Example 3 — La-Z-Boy (Retail), 10 shots. Blue: Qwen3-VL 32B selected; yellow: golden shot (ground truth).
Category
# Questions
Agreement (%)
Action
7
93.96
Information
3
74.47
Emotion
5
76.44
Imagery
3
70.63
Overall
18
81.96
Appendix
Table 11: Agreement between Gemini 3.1 Pro and AdNotator on a random sample of 1,000 AdShot shots, by content-feature category and overall. Each shot is labeled on the same 18 binary questions (7 Action, 3 Information, 5 Emotion, 3 Imagery), giving 18,000 matched question–shot decisions.
Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle-speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.
Yiran Yang, Zhaowei Liu, Yuan Yuan +10
1Kuaishou Technology · 2Beijing University of Posts and Telecommunications · 3Shandong University
In recent years, large language models (LLMs) have made rapid progress in information retrieval, yet existing research has mainly focused on text or static multimodal settings. Open-domain video shot retrieval, which involves richer temporal structure and more complex semantics, still lacks systematic benchmarks and analysis. To fill this gap, we introduce ShotFinder, a benchmark that formalizes editing requirements as keyframe-oriented shot descriptions and introduces five types of controllable single-factor constraints: Temporal order, Color, Visual style, Audio, and Resolution. We curate 1,210 high-quality samples from YouTube across 20 thematic categories, using large models for generation with human verification. Based on the benchmark, we propose ShotFinder, a text-driven three-stage retrieval and localization pipeline: (1) query expansion via video imagination, (2) candidate video retrieval with a search engine, and (3) description-guided shot localization. Experiments on multiple closed-source and open-source models reveal a significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges. These results reveal that open-domain video shot retrieval is still a critical capability that multimodal large models have yet to overcome.
Short-form videos have become a primary medium for digital advertising, requiring scalable and efficient content creation. However, current workflows and AI tools remain disjoint and modality-specific, leading to high production costs and low overall efficiency. To address this issue, we propose AutoCut, an end-to-end advertisement video editing framework based on multimodal discretization and controllable editing. AutoCut employs dedicated encoders to extract video and audio features, then applies residual vector quantization to discretize them into unified tokens aligned with textual representations, constructing a shared video-audio-text token space. Built upon a foundation model, we further develop a multimodal large language model for video editing through combined multimodal alignment and supervised fine-tuning, supporting tasks covering video selection and ordering, script generation, and background music selection within a unified editing framework. Finally, a complete production pipeline converts the predicted token sequences into deployable long video outputs. Experiments on real-world advertisement datasets show that AutoCut reduces production cost and iteration time while substantially improving consistency and controllability, paving the way for scalable video creation.