Beyond Covers and Titles: Evidence-First Ranking for Robust Cold-Start Recommendation
Authors: Fuyi Wang, Yuxin Cao, Yifan Liao, Renyang Liu, Lin Li, Wei Song
Organizations: Royal Melbourne Institute of Technology, Australia · National University of Singapore, Singapore · The Hong Kong University of Science and Technology (Guangzhou), China · Southern Cross University, Australia · Griffith University, Australia
Short-video platforms rank new uploads before anyone has watched them, using content towers that read a video's cover, title and frames. Existing attacks on content-aware recommenders perturb pixels or features and ignore platform re-encoding, and existing defenses train against norm-bounded perturbations that neither match real edits nor preserve recommendation quality. We show with REAP, a realizable attack through {a simulated platform channel (H.264 transcoding, rescaling, cropping and JPEG)}, that packaging edits raise a cold video by 32.0 rank-percentile points in unrelated audiences. We then observe that cold-start ranking does not need the packaging at all: a tower that sees only the video body ranks cold items best. Building on this, we propose Evidence-First Ranking (EFR), which scores cold videos from their body, admits packaging only through a support-gated residual bounded by γ, and merges cold and warm rankings by rank percentile, yielding an exact certificate against any packaging edit. On MicroLens-100K, EFR cuts the attack gain by 81% and raises cold-start Recall@100 from 0.1133 to 0.170, while keeping the merged NDCG@10 within 4% of the undefended system.
Figures & tables
CLIP
SigLIP
Hybrid
Random control
26.3
25.7
12.2
REAP
34.3
32.0
19.3
Over control
8.0
6.3
7.1
Other-user gain
26.2
25.6
9.3
Reach (%)
1.1 → 17.0
1.5 → 15.2
1.7 → 7.9
Table 1: Attack performance of REAP on undefended towers.
Figure 1: Pilot study of edit families and pixel attacks.
Method
NDCG@10
Cold R@100
REAP gain
Over control
Other-user
Reach (%)
Body change
Retention
Free (undefended)
0.0475
0.1133
29.2
4.1
24.6
1.6 → 12.8
9.8
1.00
Body
0.0210
0.1697
0.0
0.0
0.0
2.9 → 2.9
23.6
2.40
Cover text masking *
0.0410
0.0989
29.6
n/a
23.2
1.6 → 12.8
n/a
n/a
Random-edit invariance
0.0463
0.1152
23.3
1.2
21.1
1.4 → 6.1
9.8
1.00
AMR, ϵ=0.1
0.0304
0.1485
2.4
0.3
1.9
1.5 → 1.7
24.6
2.51
UAT-MC, ϵ=0.1
0.0413
0.1418
16.5
3.5
14.2
1.1 → 3.4
19.3
1.96
Table 2: Performance of defenses that keep reading the packaging on SigLIP.
Model
Held-out gain
Reach (%)
Honest edit
Cover
18.5
1.1 → 4.7
−8.5
Free
17.9
1.0 → 5.6
−10.3
Random-edit invariance
18.8
1.0 → 4.4
−1.8
AMR
18.9
1.0 → 3.1
−0.9
UAT-MC
19.5
1.0 → 3.1
−0.6
Strategic training
10.0
1.1 → 2.0
−6.7
Table 3: Unseen edits and honest creators on SigLIP.
Figure 2: Overview of Evidence-First Ranking (EFR). EFR scores cold videos by the body, caps packaging at γ and merges by rank.
Strategic training
Strategic + bottleneck
Variant
NDCG
Gain
Title
NDCG
Gain
Title
Full method
0.0430
19.6
3.6
0.0391
14.8
1.1
No Linv , no Limp
0.0477
29.3
1.3
0.0437
29.8
6.3
No Limp
0.0431
18.9
1.2
0.0402
16.8
-0.8
Random orbit candidate
0.0435
23.6
0.7
0.0378
18.9
1.5
Listwise KL invariance
0.0486
23.2
-0.6
0.0430
24.2
6.4
Table 4: Ablations of strategic training on SigLIP.
System
Utility
Robustness
NDCG@10
Cold R@100
REAP gain
Reach (%)
Certified gain
Certified <5 (%)
Free (undefended)
0.0469
0.1133
29.2
1.6 → 12.8
–
–
Router, body only ( γ=0 )
0.0454
0.1637
0.0 (by construction)
unchanged
0.0
100
EFR, γ=0.02
0.0451
0.1699
5.4
3.3 → 5.0
7.5
31
EFR, γ=0.05
0.0452
0.1684
10.9
2.9 → 6.1
16.4
8
EFR, γ=0.1
0.0458
0.1556
17.4
3.0 → 9.2
27.7
1
Table 5: Defense performance of EFR in the merged list on SigLIP. † Attacked on fewer seeds than were trained.
CLIP
SigLIP
MicroLens-100K
ShortVideo
MicroLens-100K (ID-hybrid)
ShortVideo
System
NDCG
Cold
Gain
Cert.
NDCG
Cold
Gain
Cert.
NDCG
Cold
Gain
Cert.
NDCG
Cold
Gain
Cert.
Free
0.0395
0.1149
33.8
–
0.0092
0.0294
31.1
–
0.0530
0.1340
20.5
–
0.0081
0.0352
31.6
–
UAT-MC
–
–
–
–
0.0088
0.0317
20.6
–
–
–
–
–
0.0070
0.0388
14.9
–
Strategic training
0.0317
0.1046
17.9
–
–
–
–
–
0.0524
0.1276
4.8
–
–
–
–
–
Body router
0.0374
0.160
0.0
0.0
0.0093
0.0252
0.0
0.0
0.0510
0.165
0.0
0.0
0.0082
0.0346
0.0
0.0
Table 6: Defense performance across encoders, warm rankers and datasets.
Figure 3: Certified gain of the attention bottleneck.
Figure 4: Adaptive attack by splicing cohort footage.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: REAP gain per target-cohort pair on SigLIP. (a) Cumulative distribution over 900 pairs. (b) Mean gain per target cohort.
Figure 6: Certified against empirical REAP gain per target-cohort pair on SigLIP.
Model
NDCG@10
Cold R@100
Pool
REAP gain
Cover
Title
Cover (undefended)
0.0400
0.1006
28.4
32.0
0.7
-2.0
Free (undefended)
0.0475
0.1133
26.2
29.2
-0.2
2.9
AMR, Cover tower, ϵ=0.02
0.0390
0.1248
27.1
30.4
0.7
-3.4
UAT-MC, Cover tower, ϵ=0.02
0.0388
0.1255
26.9
30.2
0.6
-3.5
Strategic training
0.0409
0.1135
13.7
16.4
0.2
2.5
Strategic training, λ1=1
0.0455
0.1204
14.2
17.2
0.3
2.0
Appendix
Table 7: Further baselines and attack diagnostics on SigLIP.
Figure 7: Edit operations in the programs selected by REAP.
We propose Compressed Video Aggregator (CVA), a lightweight micro-video recommendation module that decouples video information from preference learning. It aggregates frozen VFM embeddings, and uses latent reasoning without cross-attention projection, producing compact video embeddings for recommenders. Due to the redundancy in the frame count of the original benchmark and its overly coarse sampling, we used titles to re-select key frames based on CLIP. Experiments on MicroLens and Short-Video show consistent gains with orders-of-magnitude reductions in training time and GPU memory, and re-selected frames can further enhance the performance of all methods, including CVA. Furthermore, we also discussed the impact of several scenarios involving erroneous titles on our method. Code will be released soon.
Yang Xiao, Huiyuan Chen, Kaiyuan Deng +5
University of Tulsa · Amazon · University of Arizona +1
A video benchmark should reward the capability it claims to measure, yet models can exploit answer options, question text, or partial visual evidence. We introduce the attack pyramid, five levels of shortcut attacks with increasing access to each item, and audit 115 video benchmarks with it. On 35 benchmarks, attackers that never see a frame approach full-video accuracy. On 51 benchmarks with temporal probes, shuffled frames keep a median 96% of full-video accuracy. Near-duplicate questions make up at least half the items in 63 benchmarks. We screen 505,518 question-answer pairs from 112 of them into an audited pool. Agents turn evaluation requests into specifications, and a deterministic selector with a red-team gate composes reproducible benchmarks. We release Video-Index, the 210 hardest verified items under these attacks in each of four capability groups, 840 items from 76 sources. With the same fixed input, Claude Opus 5 outscores every open-source model by over 37 percentage points, and agent tools add about 20 more, yet all systems leave room to improve efficiency and accuracy. Blog: https://www.enxinsong.com/blog/video-index/ GitHub: https://github.com/Espere-1119-Song/Video-Index Hugging Face: https://huggingface.co/datasets/Video-Index/Video-Index
Enxin Song, Yinuo Xu, Shusheng Yang +2
University of Pennsylvania · New York University · Princeton University
Personalized video recommendation predicts user preference at the video level, while temporal video grounding localizes query-relevant moments. However, strong localization does not establish whether the retrieved moment constitutes valid evidence for recommending the video to a particular user. We study counterfactual behavior-grounded evidence retrieval, which separates where personalized evidence occurs from whether such evidence exists and evaluates whether model predictions respond consistently when that evidence is replaced. We introduce CBGER-10K, containing 5,000 controlled factual--counterfactual pairs for 3,026 users, where each pair replaces only the focal behavior-supported segment while preserving the user, temporal position, and hard distractors. We further propose CBGER, a compact framework that decouples segment-level localization from video-level evidence estimation and learns both through structured counterfactual supervision. CBGER achieves 0.4432 MRR, 0.6977 Pair Accuracy, and 0.6987 Intervention Consistency across five adapted personalized-highlight and temporal-grounding baselines. Notably, compared with QD-DETR, its MRR improvement is not statistically significant, while Pair Accuracy improves by 11.03 points. These results show that accurate temporal localization does not necessarily imply reliable personalized evidence existence, motivating explicit evaluation of Whether alongside Where.