Pith. sign in

REVIEW 3 major objections 5 minor 59 references

A benchmark for short-video engagement prediction shows large multimodal models outperform a feature-based baseline, with the top entry reaching SROCC 0.707.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-05 11:14 UTC pith:ERRWXUN5

load-bearing objection A useful engagement-prediction benchmark whose leaderboard numbers are probably trustworthy except that the undisclosed ECR label transform leaves PLCC uninterpretable. the 3 major comments →

arxiv 2509.02969 v1 pith:ERRWXUN5 submitted 2025-09-03 cs.CV cs.MMcs.SI

VQualA 2025 Challenge on Engagement Prediction for Short Videos: Methods and Results

Dasong Li , Sizhuo Ma , Hang Hua , Wenjie Li , Jian Wang , Chris Wei Zhou , Fengbin Guan , Xin Li
show 35 more authors
Zihao Yu Yiting Lu Ru-Ling Liao Yan Ye Zhibo Chen Wei Sun Linhan Cao Yuqin Cao Weixia Zhang Wen Wen Kaiwei Zhang Zijian Chen Fangfang Lu Xiongkuo Min Guangtao Zhai Erjia Xiao Lingfeng Zhang Zhenjie Su Hao Cheng Yu Liu Renjing Xu Long Chen Xiaoshuai Hao Zhenpeng Zeng Jianqin Wu Xuxu Wang Qian Yu Bo Hu Weiwei Wang Pinxin Liu Yunlong Tang Luchuan Song Jinxi He Jiaru Wu Hanjia Lyu
This is my paper
classification cs.CV cs.MMcs.SI
keywords engagement predictionshort-form videoengagement continuation rateSnapUGClarge multimodal modelsvideo quality assessmentbenchmarkcold-start recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper reports the design and results of a community challenge on predicting how engaging a short video will be, using a new large dataset of real user interactions from a social media platform. The task is to predict the engagement continuation rate, the probability that a viewer watches past the first 5 seconds, using the video, its audio, title, and description. The challenge shows that the best methods are large multimodal models: an ensemble of two video-language models reaches SROCC 0.707 and PLCC 0.714, clearly outperforming the provided multi-modal feature baseline (SROCC 0.657, PLCC 0.665). The paper also shows that engagement prediction is not the same as video quality assessment, and that even text-only inputs carry real predictive signal. If the benchmark is sound, it gives content platforms and creators a practical way to estimate cold-start popularity before a video has accumulated many views.

Core claim

The central claim is that engagement continuation rate (ECR) — the probability that a viewer watches beyond 5 seconds — can be predicted from video content better with large multimodal foundation models than with hand-crafted multi-modal features. The organizers built a 120,651-video dataset from public short videos with engagement labels aggregated from real viewer behavior, then ran a challenge in which 15 teams submitted test predictions. The winning entry combined two large multimodal models and reached SROCC 0.707 and PLCC 0.714, a meaningful gain over the provided baseline; the authors attribute the gains mainly to large multimodal models, while noting that a text-only gradient-boosted

What carries the argument

The central objects are two engagement metrics defined on the SnapUGC dataset: normalized average watch percentage (NAWP), which rescales average watch time by duration, and engagement continuation rate (ECR), defined as P(watch > 5s), which is duration-stable and used as the challenge target. For the challenge, the raw ECR is replaced by a rank-normalized version to protect creator privacy. The comparison that carries the argument is between the provided multi-modal baseline and the submitted methods, especially the winning ensemble of Video-LLaMA2 and Qwen2.5-VL fine-tuned to regress ECR from sampled frames, audio, and text.

Load-bearing premise

The leaderboard values are valid only if the rank-normalized ECR challenge labels preserve the ordering of the raw engagement continuation rate; the normalization is not described, so a non-monotonic transform would make the reported SROCC/PLCC measure something other than the stated task.

What would settle it

Re-score the submitted models or a sample of test videos against the raw ECR values P(watch > 5s) instead of the rank-normalized labels; if the Spearman correlation changes materially from the reported 0.707, the leaderboard is an artifact of the undisclosed normalization.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the benchmark is sound, existing no-reference video quality models trained on small labeled sets are not enough for engagement prediction; platforms need models trained on interaction-derived labels at scale.
  • Large multimodal models fine-tuned for regression on engagement labels become the state of practice for this task, with the winning ensemble demonstrating that combining a video-language model with a general vision-language model helps.
  • Text and audio metadata carry signal even without visual content, so multimodal fusion should treat title, description, and sound classification as first-class inputs, not auxiliary cues.
  • ECR as a target is well suited to short video: it isolates the first 5 seconds, is duration-independent, and correlates strongly (0.928) with the normalized watch-time metric NAWP.
  • A benchmark of this size gives researchers and platforms a common testbed for cold-start popularity prediction, allowing direct comparison of engagement models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the challenge labels are rank-normalized ECR values and the normalization procedure is not described, the reported SROCC/PLCC could depend on the transform; a non-monotonic normalization would mean leaderboard scores do not measure prediction of raw retention.
  • The dataset only includes videos with more than 2,000 views, so the paper's cold-start motivation is not directly tested; the true cold-start setting (videos with very few initial interactions) would be a natural follow-up benchmark.
  • The strong showings of large multimodal models suggest that engagement prediction could be improved further by prompt-based reasoning or by models that ingest the first few seconds at native temporal resolution, given the ECR definition.
  • If the rank-normalized labels are later released with their formula, the same submitted models could be re-scored against the raw ECR, providing a clean check on how much of the measured performance is due to the normalization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper reports the organization and results of the VQualA 2025 Challenge on Engagement Prediction for Short Videos, held at ICCV 2025. The challenge uses the SnapUGC dataset of 120,651 short videos, with engagement labels derived from real user interactions. The core prediction target is the Engagement Continuation Rate (ECR), defined as P(watch > 5s), with a rank-normalized version used for the challenge. The paper describes the dataset, the engagement metrics (NAWP and ECR), a multi-modal baseline built on prior work by Li et al. [28], and the methods of seven participating teams. The main empirical claim is that the best team (ECNU-SJTU VQA) achieves SROCC 0.707 and PLCC 0.714, outperforming the baseline (SROCC 0.657, PLCC 0.665), and that large multimodal models were the main driver of progress. The paper also highlights that a text-only XGBoost model (brucelyu17) achieves competitive results.

Significance. If the reported results are reliable, this challenge provides a useful benchmark for engagement prediction from video content, with the notable strength of using real-world interaction data from over 2,000 users rather than small-scale subjective annotation. The paper is transparent in publishing participant methods and the baseline, and it contributes to the growing evidence that large multimodal models transfer to engagement-related tasks. The challenge also documents a meaningful gap between traditional VQA-based features and LMM-based approaches. However, the central leaderboard claim depends on the exact definition of the challenge labels, and that definition is not fully specified in the manuscript.

major comments (3)
  1. [Section 2.2, 'ECR for the Challenge'] The paper states that 'the ECR used in this challenge is derived from normalizing the ranking of real ECR,' but it does not specify the normalization transformation. This is load-bearing because Table 2 reports PLCC values, and PLCC is not invariant under arbitrary monotonic transformations. If the normalization is nonlinear, the reported PLCC (e.g., 0.714 for ECNU-SJTU VQA) does not measure Pearson correlation with the true ECR = P(watch > 5s) described in the task definition. Moreover, teams are prompted to output a probability in [0,1] (Figure 4), while the training labels may be rank-percentiles, creating a distribution mismatch. The authors should provide the exact normalization formula, state whether it is monotonic, clarify whether it is fitted on the training set or the full dataset (to rule out test-set leakage), and, if privacy permits, recompute PLCC against raw ECR values or
  2. [Section 1 vs. Section 2.1] The introduction motivates the challenge primarily with the cold-start problem, describing videos with about one hundred initial views and limited user reactions. However, Section 2.1 states that videos with view counts exceeding 2,000 are selected, and no cold-start setting is evaluated. The central claim that this benchmark supports cold-start engagement prediction is therefore not supported by the dataset construction. The authors should either reframe the motivation as general engagement prediction for established videos, or add experiments on videos with fewer views to justify the cold-start connection.
  3. [Table 2] The table reports a single SROCC/PLCC value per team with no error bars, confidence intervals, or significance tests. The top two teams differ by only 0.012 in final score, and the difference between teams 1 and 3 is 0.030. Since the challenge's headline claim is that LMM-based methods significantly outperform the baseline and each other, the absence of variance or repeated-run information makes it difficult to assess whether these differences are meaningful. The authors should report standard deviations over multiple runs or at least a statistical comparison between the top submission and the baseline.
minor comments (5)
  1. [Section 2.2, Eq. (1)] The fitted parameters alpha and beta for fmax(d) are not reported. If NAWP is used only as motivation for ECR, this should be stated; otherwise, provide the fitted values or a reference to the dataset paper.
  2. [Table 2] There are naming inconsistencies: 'ICML-DAMO' appears in the table but 'IMCL-DAMO' is used elsewhere, and the row 'MCCE (MCCE' contains a duplicate. Also, 'EasyVQA' is formatted with an extra space in the table.
  3. [Section 2.1] Typo: 'Datasset' should be 'Dataset.' In Figure 2, 'A WT' should be 'AWT.'
  4. [Section 4.2] Typo: 'LaV A' should be 'LLaVA.' Also, the text says 'We evaluate our framework' in a team section; this should be 'they evaluate' for consistency.
  5. [Section 3] The paper mentions 15 valid test submissions but lists only 7 teams in Table 2. Clarify whether the remaining submissions are omitted for lack of fact sheets or for other reasons.

Circularity Check

0 steps flagged

No significant circularity; the challenge is a self-contained benchmark evaluation.

full rationale

The paper's derivation chain is: real user watch-time events define ECR = P(watch > 5s); for privacy the challenge labels are a rank-normalized version of that ECR; participants train models on these labels and are scored by SROCC/PLCC on a held-out test set. The reported leaderboard numbers are new empirical measurements, not derived from the label definition by construction. The NAWP metric is indeed fitted to the top-3% linear trend in the data, but NAWP is not the challenge target; ECR is selected based on the observed 0.928 correlation, which is an empirical model-selection step, not circular. The paper's reliance on [28] for the SnapUGC dataset and the baseline is a normal self-citation; [28] is prior peer-reviewed work with an externally available dataset, and the challenge results are self-contained against that baseline. The unreported exact normalization of ECR raises reproducibility and PLCC-interpretation concerns, but that is a data-validity or reporting issue, not a circular equivalence between the prediction target and the model output.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper's proposed metrics NAWP and ECR are empirical constructs: NAWP depends on fitted alpha/beta and the top-3% choice, and the challenge target is a privatized rank-normalized version of ECR whose exact form is not given. No new physical entities are introduced.

free parameters (4)
  • alpha (slope) in fmax(d) = not reported
    Fitted to the top 3% of average watch times as a linear function of duration (Eq. 1); the paper states it may vary across datasets.
  • beta (intercept) in fmax(d) = not reported
    Same fit as alpha; used to define NAWP normalization.
  • top 3% AWT threshold = 3%
    Arbitrary percentile chosen to define 'most popular' videos in NAWP construction.
  • ECR 5-second threshold = 5 s
    The definition of ECR uses P(watch > 5s); this is a design choice, not fitted to data.
axioms (4)
  • domain assumption Videos with top 3% of highest AWT, regardless of duration, are equally most popular
    Used to fit fmax(d) and define NAWP in Eq. 1-2; not independently validated.
  • domain assumption Minimum average watch time of 0 seconds marks the least popular videos
    Sets fmin(d)=0 in Eq. 1; plausible but unstated scaling assumption.
  • domain assumption ECR and NAWP have a strong correlation (0.928) that justifies choosing ECR as the challenge metric
    Empirical observation from the dataset; the correlation value is reported without confidence intervals.
  • ad hoc to paper The rank-normalized ECR labels preserve the engagement semantics of the raw ECR
    Required for the challenge scores to reflect the defined metric, but the normalization is not specified.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 14057 in / 13559 out tokens · 122580 ms · 2026-08-05T11:14:00.340682+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of VQualA 2025 Challenge on Engagement Prediction for Short Videos: Methods and Results." pith.science (2026). https://pith.science/paper/ERRWXUN5

@misc{pith2026250902969,
  author       = {Pith},
  title        = {Pith review of: VQualA 2025 Challenge on Engagement Prediction for Short Videos: Methods and Results},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERRWXUN5}},
  note         = {Machine review of arXiv:2509.02969}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents an overview of the VQualA 2025 Challenge on Engagement Prediction for Short Videos, held in conjunction with ICCV 2025. The challenge focuses on understanding and modeling the popularity of user-generated content (UGC) short videos on social media platforms. To support this goal, the challenge uses a new short-form UGC dataset featuring engagement metrics derived from real-world user interactions. This objective of the Challenge is to promote robust modeling strategies that capture the complex factors influencing user engagement. Participants explored a variety of multi-modal features, including visual content, audio, and metadata provided by creators. The challenge attracted 97 participants and received 15 valid test submissions, contributing significantly to progress in short-form UGC video engagement prediction.

Figures

Figures reproduced from arXiv: 2509.02969 by Bo Hu, Chris Wei Zhou, Dasong Li, Erjia Xiao, Fangfang Lu, Fengbin Guan, Guangtao Zhai, Hang Hua, Hanjia Lyu, Hao Cheng, Jianqin Wu, Jian Wang, Jiaru Wu, Jinxi He, Kaiwei Zhang, Lingfeng Zhang, Linhan Cao, Long Chen, Luchuan Song, Pinxin Liu, Qian Yu, Renjing Xu, Ru-Ling Liao, Sizhuo Ma, Wei Sun, Weiwei Wang, Weixia Zhang, Wenjie Li, Wen Wen, Xiaoshuai Hao, Xin Li, Xiongkuo Min, Xuxu Wang, Yan Ye, Yiting Lu, Yu Liu, Yunlong Tang, Yuqin Cao, Zhenjie Su, Zhenpeng Zeng, Zhibo Chen, Zihao Yu, Zijian Chen.

Figure 1
Figure 1. Figure 1: Sample frames of the short videos in SnapUGC dataset [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a), (d): The distributions of average watch time (AWT) and engagement continuation rate (ECR), respectively. ECR, calculated [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overview framework of baseline. LLM decoder Visual projector Visual encoder : The video frames:<frames tokens><frames tokens> … …<frames tokens>, the audio :<audio tokens>: how would you judge the engagement continuation rate of the given content, where engagement continuation rate represents the probability of watch time exceeding 5 seconds. The title of the video is :<title>, and the description of t… view at source ↗
Figure 4
Figure 4. Figure 4: The overview framework provided by Team ECNU-SJTU VQA. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The overview framework provided by Team IMCL-DAMO. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The overview framework provided by Team HKUST-Cardiff-MI-BAAI . [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The overview framework provided by Team EasyVQA. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 49 canonical work pages · 4 internal anchors

  1. [1]

    Mart ´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Man ´e, Rajat Monga, Sherry Moore, Derek M...

    work page 2015

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 5, 6

    Pith/arXiv arXiv 2025

  3. [3]

    VLEngagement: A Dataset of Scientific Video Lectures for Evaluating Population-based Engagement

    Sahan Bulathwela, Maria Perez-Ortiz, Emine Yilmaz, and John Shawe-Taylor. VLEngagement: A Dataset of Sci- entific Video Lectures for Evaluating Population-based En- gagement. arXiv e-prints, art. arXiv:2011.02273, 2020. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  4. [4]

    Learning generalized spatial-temporal deep feature representation for no-reference video quality as- sessment

    Baoliang Chen, Lingyu Zhu, Guo Li, Fangbo Lu, Hongfei Fan, and Shiqi Wang. Learning generalized spatial-temporal deep feature representation for no-reference video quality as- sessment. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):1903–1916, 2022. 2

    work page 1903

  5. [5]

    Rirnet: Recurrent-in-recurrent network for video quality assessment

    Pengfei Chen, Leida Li, Lei Ma, Jinjian Wu, and Guangming Shi. Rirnet: Recurrent-in-recurrent network for video quality assessment. In Proceedings of the 28th ACM International Conference on Multimedia , page 834–842, New York, NY , USA, 2020. Association for Computing Machinery. 2

    work page 2020

  6. [6]

    Beats: Au- dio pre-training with acoustic tokenizers

    Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Au- dio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022. 7

    Pith/arXiv arXiv 2022

  7. [7]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016. 8

    work page 2016

  8. [8]

    DiffVQA: Video Quality Assessment Using Diffusion Feature Extractor

    Wei-Ting Chen, Yu-Jiet V ong, Yi-Tsung Lee, Sy-Yen Kuo, Qiang Gao, Sizhuo Ma, and Jian Wang. Diffvqa: Video quality assessment using diffusion feature extractor. arXiv preprint arXiv:2505.03261, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Vquala 2025 challenge on genai-bench aigc video quality assessment: Methods and re- sults

    Ying Chen, Huasheng Wang, Pengxiang Xiao, Yukang Ding, Enpeng Liu, Wei Zhou, and et al. Vquala 2025 challenge on genai-bench aigc video quality assessment: Methods and re- sults. In Proceedings of the IEEE/CVF Conference on Com- puter Vision (ICCV) Workshops, pages 1–11, 2025. 3

    work page 2025

  10. [11]

    Videollama 2: Advancing spatial- temporal modeling and audio understanding in video-llms

    Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial- temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 5, 6

    Pith/arXiv arXiv 2024

  11. [12]

    Imagenet: A large-scale hierarchical im- age database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical im- age database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2

    work page 2009

  12. [13]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the asso- ciation for computational linguistics: human language tech- nologies, volume 1 (long and short papers) , pages 4171– 4186, 2019. 7

    work page 2019

  13. [14]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7

    Pith/arXiv arXiv 2010

  14. [15]

    Konvid-150k: A dataset for no-reference video qual- ity assessment of videos in-the-wild

    Franz G ¨otz-Hahn, Vlad Hosu, Hanhe Lin, and Dietmar Saupe. Konvid-150k: A dataset for no-reference video qual- ity assessment of videos in-the-wild. In IEEE Access 9 , pages 72139–72160. IEEE, 2021. 2

    work page 2021

  15. [16]

    InternVQA: Advancing Compressed Video Quality Assessment with Distilling Large Foundation Model

    Fengbin Guan, Zihao Yu, Yiting Lu, Xin Li, and Zhibo Chen. Internvqa: Advancing compressed video qualityassessment with distilling large foundation model. arXiv preprint arXiv:2502.19026, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  16. [17]

    Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 6546– 6555, 2018. 2

    work page 2018

  17. [18]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 2

    work page 2016

  18. [19]

    The konstanz natural video database (konvid-1k)

    Vlad Hosu, Franz Hahn, Mohsen Jenadeleh, Hanhe Lin, Hui Men, Tam´as Szir ´anyi, Shujun Li, and Dietmar Saupe. The konstanz natural video database (konvid-1k). In Ninth Inter- national Conference on Quality of Multimedia Experience (QoMEX), pages 1–6, 2017. 2

    work page 2017

  19. [20]

    Vquala 2025 doc- ument image quality assessment challenge

    Fan Huang, Xiongkuo Min, Zhichao Ma, Xiaohong Liu, Chris Wei Zhou, Guangtao Zhai, and et al. Vquala 2025 doc- ument image quality assessment challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision (ICCV) Workshops, pages 1–8, 2025. 3

    work page 2025

  20. [21]

    The kinetics human action video dataset

    Will Kay, Jo ˜ao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Apostol Natsev, Mustafa Suley- man, and Andrew Zisserman. The kinetics human action video dataset. ArXiv, abs/1705.06950, 2017. 2

    Pith/arXiv arXiv 2017

  21. [22]

    Guo, Daniel T

    Juho Kim, Philip J. Guo, Daniel T. Seaton, Piotr Mitros, Krzysztof Z. Gajos, and Robert C. Miller. Understanding in- video dropouts and interaction peaks inonline lecture videos. In Proceedings of the First ACM Conference on Learning @ 9 Scale Conference, page 31–40, New York, NY , USA, 2014. Association for Computing Machinery. 4

    work page 2014

  22. [23]

    Melu: Meta-learned user preference esti- mator for cold-start recommendation

    Hoyeop Lee, Jinbae Im, Seongwon Jang, Hyunsouk Cho, and Sehee Chung. Melu: Meta-learned user preference esti- mator for cold-start recommendation. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, page 1073–1082, New York, NY , USA, 2019. Association for Computing Machinery. 1

    work page 2019

  23. [24]

    Quality assessment of in-the-wild videos

    Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In Proceedings of the 27th ACM International Conference on Multimedia , page 2351–2359, New York, NY , USA, 2019. Association for Computing Machinery. 2

    work page 2019

  24. [25]

    Learning degradation rep- resentations for image deblurring

    Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation rep- resentations for image deblurring. In Computer Vision – ECCV 2022, pages 736–753, Cham, 2022. Springer Nature Switzerland. 5

    work page 2022

  25. [26]

    Efficient burst raw denois- ing with variance stabilization and multi-frequency denois- ing network, 2022

    Dasong Li, Yi Zhang, Ka Lung Law, Xiaogang Wang, Hong- wei Qin, and Hongsheng Li. Efficient burst raw denois- ing with variance stabilization and multi-frequency denois- ing network, 2022

    work page 2022

  26. [27]

    A simple baseline for video restoration with grouped spatial- temporal shift

    Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A simple baseline for video restoration with grouped spatial- temporal shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9822–9832, 2023. 5

    work page 2023

  27. [28]

    Delving deep into engagement prediction of short videos

    Dasong Li, Wenjie Li, Baili Lu, Hongsheng Li, Sizhuo Ma, Gurunandan Krishnan, and Jian Wang. Delving deep into engagement prediction of short videos. In Computer Vision – ECCV 2024, pages 289–306, Cham, 2025. Springer Nature Switzerland. 2, 4, 5

    work page 2024

  28. [29]

    Vquala 2025 challenge on image super-resolution generated content qual- ity assessment: Methods and results

    Yixiao Li, Xin Li, Wei Zhou, Shuo Xing, Hadi Amirpour, Xiaoshuai Hao, Guanghui Yue, Baoquan Zhao, Weide Liu, Xiaoyuan Yang, Zhengzhong Tu, and et al. Vquala 2025 challenge on image super-resolution generated content qual- ity assessment: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision (ICCV) Work- shops, pages 1–10, 2025. 3

    work page 2025

  29. [30]

    Efficient video quality assessment with deeper spatiotemporal feature extraction and integra- tion

    Yinhao Liu, Xiaofei Zhou, Haibing Yin, Hongkui Wang, and Chenggang Yan. Efficient video quality assessment with deeper spatiotemporal feature extraction and integra- tion. Journal of Electronic Imaging, 30:063034, 2021. 2

    work page 2021

  30. [31]

    Q-adapt: Adapting lmm for visual quality as- sessment with progressive instruction tuning

    Yiting Lu, Xin Li, Haoning Wu, Bingchen Li, Weisi Lin, and Zhibo Chen. Q-adapt: Adapting lmm for visual quality as- sessment with progressive instruction tuning. arXiv preprint arXiv:2504.01655, 2025. 2

    Pith/arXiv arXiv 2025

  31. [32]

    Reduced-reference video quality assessment of compressed video sequences

    Lin Ma, Songnan Li, and King Ngi Ngan. Reduced-reference video quality assessment of compressed video sequences. IEEE Transactions on circuits and systems for video tech- nology, 22(10):1441–1456, 2012. 2

    work page 2012

  32. [33]

    Vquala 2025 challenge on face image quality assessment: Methods and results

    Sizhuo Ma, Wei-Ting Chen, Qiang Gao, Jian Wang, Chris Wei Zhou, Wei Sun, Weixia Zhang, Linhan Cao, Jun Jia, Xiangyang Zhu, Dandan Zhu, Xiongkuo Min, Guang- tao Zhai, Baoying Chen, Xiongwei Xiao, Jishen Zeng, Wei Wu, Tiexuan Lou, Yuchen Tan, Chunyi Song, Zhiwei Xu, MohammadAli Hamidi, Hadi Amirpour, Mingyin Bai, Ji- awang Du, Zhenyu Jiang, Zilong Lu, Zigua...

    work page 2025

  33. [34]

    Warm up cold-start advertisements: Improving ctr predictions via learning to learn id embeddings

    Feiyang Pan, Shuokai Li, Xiang Ao, Pingzhong Tang, and Qing He. Warm up cold-start advertisements: Improving ctr predictions via learning to learn id embeddings. In Proceed- ings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval , page 695–704, New York, NY , USA, 2019. Association for Com- puting Machinery. 1

    work page 2019

  34. [35]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, pages 8748–8763. PmLR, 2021. 7

    work page 2021

  35. [36]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 8

    Pith/arXiv arXiv 1908

  36. [37]

    Large-scale study of perceptual video quality

    Zeina Sinno and Alan Conrad Bovik. Large-scale study of perceptual video quality. IEEE Transactions on Image Pro- cessing, 28(2):612–627, 2019. 2

    work page 2019

  37. [38]

    Video quality as- sessment by reduced reference spatio-temporal entropic dif- ferencing

    Rajiv Soundararajan and Alan C Bovik. Video quality as- sessment by reduced reference spatio-temporal entropic dif- ferencing. IEEE Transactions on Circuits and Systems for Video Technology, 23(4):684–694, 2012. 2

    work page 2012

  38. [39]

    Deep learning based full-reference and no-reference quality assessment models for compressed ugc videos

    Wei Sun, Tao Wang, Xiongkuo Min, Fuwang Yi, and Guangtao Zhai. Deep learning based full-reference and no-reference quality assessment models for compressed ugc videos. In 2021 IEEE International Conference on Multime- dia & Expo Workshops (ICMEW) , pages 1–6. IEEE, 2021. 1

    work page 2021

  39. [40]

    A deep learning based no-reference quality assessment model for ugc videos

    Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model for ugc videos. In Proceedings of the 30th ACM Interna- tional Conference on Multimedia, pages 856–865, 2022. 2

    work page 2022

  40. [41]

    Engagement prediction of short videos with large multimodal models

    Wei Sun, Linhan Cao, Yuqin Cao, Weixia Zhang, Wen Wen, Kaiwei Zhang, Zijian Chen, Fangfang Lu, Xiongkuo Min, and Guangtao Zhai. Engagement prediction of short videos with large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision (ICCV) Work- shops, 2025. 5

    work page 2025

  41. [42]

    Mingxing Tan and Quoc V . Le. Efficientnetv2: Smaller mod- els and faster training. In Proceedings of the 38th Interna- tional Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 10096–10106. PMLR, 2021. 2, 5 10

    work page 2021

  42. [43]

    Ugc-vqa: Benchmarking blind video quality assessment for user generated content

    Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Trans- actions on Image Processing, 30:4449–4464, 2021. 2

    work page 2021

  43. [44]

    Dropoutnet: Addressing cold start in recommender systems

    Maksims V olkovs, Guangwei Yu, and Tomi Poutanen. Dropoutnet: Addressing cold start in recommender systems. In Proceedings of the 31st International Conference on Neu- ral Information Processing Systems , page 4964–4973, Red Hook, NY , USA, 2017. Curran Associates Inc. 1

    work page 2017

  44. [45]

    Skywork-vl re- ward: An effective reward model for multimodal understand- ing and reasoning, 2025

    Xiaokun Wang, Peiyu Wang, Jiangbo Pei, Wei Shen, Yi Peng, Yunzhuo Hao, Weijie Qiu, Ai Jian, Tianyidan Xie, Xuchen Song, Yang Liu, and Yahui Zhou. Skywork-vl re- ward: An effective reward model for multimodal understand- ing and reasoning, 2025. 8

    work page 2025

  45. [46]

    Rich features for perceptual quality assessment of ugc videos

    Yilin Wang, Junjie Ke, Hossein Talebi, Joong Gon Yim, Neil Birkbeck, Balu Adsumilli, Peyman Milanfar, and Feng Yang. Rich features for perceptual quality assessment of ugc videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13435–13444, 2021. 2, 5

    work page 2021

  46. [47]

    Exploring video quality assessment on user gener- ated contents from aesthetic and technical perspectives

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jing- wen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user gener- ated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20144–20154, 2023. 2, 6

    work page 2023

  47. [48]

    Q-align: Teaching lmms for visual scoring via discrete text-defined levels

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtao Zhai, and Weisi Lin. Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. Equal Contribution by Wu, Haon- ing and Zhang, Zicheng. Corresponding Au...

    Pith/arXiv arXiv 2023

  48. [49]

    Q-instruct: Improving low-level visual abilities for multi-modality foundation models

    Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 25490–25500, 2024. 2

    work page 2024

  49. [50]

    Be- yond views: Measuring and predicting engagement in online videos

    Siqi Wu, Marian-Andrei Rizoiu, and Lexing Xie. Be- yond views: Measuring and predicting engagement in online videos. Proceedings of the International AAAI Conference on Web and Social Media, 12(1), 2018. 2

    work page 2018

  50. [51]

    mplug-2: A modularized multi-modal foundation model across text, image and video

    Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jin- gren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. ArXiv, abs/2302.00402,

    Pith/arXiv arXiv

  51. [52]

    Subjective quality assessment for youtube ugc dataset

    Joong Gon Yim, Yilin Wang, Neil Birkbeck, and Balu Adsumilli. Subjective quality assessment for youtube ugc dataset. In 2020 IEEE International Conference on Image Processing (ICIP), pages 131–135, 2020. 2

    work page 2020

  52. [53]

    Patch-vq: ’patching up’ the video quality problem

    Zhenqiang Ying, Maniratnam Mandal, Deepti Ghadiyaram, and Alan Bovik. Patch-vq: ’patching up’ the video quality problem. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)) , pages 14019–14029, 2021. 2

    work page 2021

  53. [54]

    Video Quality Assessment Based on Swin TransformerV2 and Coarse to Fine Strategy

    Zihao Yu, Fengbin Guan, Yiting Lu, Xin Li, and Zhibo Chen. Video quality assessment based on swin transformerv2 and coarse to fine strategy. arXiv preprint arXiv:2401.08522 ,

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [55]

    Deconfounding duration bias in watch-time pre- diction for video recommendation

    Ruohan Zhan, Changhua Pei, Qiang Su, Jianfeng Wen, Xueliang Wang, Guanyu Mu, Dong Zheng, Peng Jiang, and Kun Gai. Deconfounding duration bias in watch-time pre- diction for video recommendation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , page 4472–4481, New York, NY , USA,

    work page

  55. [56]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 2

    work page 2018

  56. [57]

    Md-vqa: Multi-dimensional quality assessment for ugc live videos

    Zicheng Zhang, Wei Wu, Wei Sun, Danyang Tu, Wei Lu, Xiongkuo Min, Ying Chen, and Guangtao Zhai. Md-vqa: Multi-dimensional quality assessment for ugc live videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 1746–1755,

    work page

  57. [58]

    Vquala 2025 challenge on visual quality comparison for large multimodal models: Methods and results

    Hanwei Zhu, Haoning Wu, Zicheng Zhang, Lingyu Zhu, Yixuan Li, Peilin Chen, Shiqi Wang, Wei Zhou, Linhan Cao, Wei Sun, Xiangyang Zhu, Weixia Zhang, Yucheng Zhu, Jing Liu, Dandan Zhu, Guantao Zhai, Xiongkuo Min, Zhichao Zhang, Xinyue Li, Shubo Xu, Anh Dao, Yifan Li, Hongyuan Yu, Jiaojiao Yi, Yiding Tian, Yupeng Wu, Feiran Sun, Jiao Lijuan, and Song Jiang. V...

    work page 2025

  58. [59]

    Learning to warm up cold item embeddings for cold-start recommenda- tion with meta scaling and shifting networks

    Yongchun Zhu, Ruobing Xie, Fuzhen Zhuang, Kaikai Ge, Ying Sun, Xu Zhang, Leyu Lin, and Juan Cao. Learning to warm up cold item embeddings for cold-start recommenda- tion with meta scaling and shifting networks. In Proceed- ings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , page 1167–1176, New York, N...

    work page 2021

  59. [2022]

    Association for Computing Machinery. 2