Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

A production livestream moderation system that pairs supervised classification with reference-based similarity matching, distilling a multimodal LLM into lightweight models, reports a 6–8% reduction in user views of unwanted streams.

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

2026-08-03 18:44 UTC pith:NTKZ6B27

load-bearing objection A credible production-system report with a real but unverified headline number; the abstract overstates what the tables show, and the 22% complementarity claim has no supporting measurement. the 5 major comments →

arxiv 2512.03553 v3 pith:NTKZ6B27 submitted 2025-12-03 cs.CV cs.AI

Dynamic Content Moderation in Livestreams: Combining Supervised Classification with MLLM-Boosted Similarity Matching

Wei Chee Yew , Hailun Xu , Sanjay Saha , Xiaotian Fan , Hiok Hian Ong , David Yuchen Wang , Kanchan Sarkar , Zhenheng Yang
show 1 more author
Danhui Guan
This is my paper
classification cs.CV cs.AI
keywords content moderationlivestreammultimodalknowledge distillationsimilarity matchingsupervised classificationvideo retrievalMLLM
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.

The paper argues that no single moderation approach suffices for livestreams: known violations are best caught by supervised multiclass classification, while novel or subtle violations require comparing each incoming clip against a stored index of verified violations. To stay within real-time latency budgets, the system distills a multimodal large language model into two lightweight students—a classifier and a re-ranking model—so the heavy reasoning happens offline. Deployed at scale, the classification pipeline reaches 67% recall at 80% precision, the similarity pipeline 76% recall at 80% precision, and online A/B tests attribute a 6–8% reduction in user views of unwanted livestreams to the combined system.

Core claim

On its own terms, the paper's central claim is that reference-based similarity matching contributes coverage that supervised classification misses—roughly 22% of detections come only from the similarity path—and that knowledge distillation from a multimodal LLM lets compact student models approach teacher-level accuracy while meeting real-time latency (about 0.4 seconds per 20-second clip for classification, about 4 seconds for reference matching). The similarity path's ability to catch novel violations rests on retrieval over a curated index of known violations, reranked by a multimodal cross-attention model; the paper reports this design detects subtle or emergent violations that evade the

What carries the argument

The central mechanism is the dual-path pipeline processing the same 20-second multimodal clip (frames, audio, ASR text). Path one is a supervised multiclass classifier (vision transformer plus audio encoder plus fusion) trained with cross-entropy and distilled from an MLLM teacher via KL and MSE losses. Path two retrieves candidates from HNSW indices using embeddings trained with MoCo-style contrastive learning plus CLIP cross-modal alignment, then reranks them with a small multimodal cross-attention model also distilled from the teacher. A clip-match aggregation step enforces temporal consistency, and sliding-window aggregation boosts re-ranking precision.

Load-bearing premise

The similarity path's ability to catch novel violations rests on the assumption that the curated index of historical violations will contain a near neighbor for any new violation type; if a new violation shares no semantic or perceptual similarity with anything already indexed, both retrieval and reranking fail.

What would settle it

Construct a held-out set of violation types collected after the reference index was built, and measure recall of the similarity path on it; if recall falls to near zero for violations that have no near neighbor in the index, the claim that similarity matching generalizes to novel violations is falsified.

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

If this is right

  • Reference matching adds roughly 22% detection coverage beyond classification alone, so hybrid systems can catch violations that a preset classifier misses.
  • Distillation from a large MLLM teacher recovers most of the accuracy gap of lightweight models (e.g., average precision from 71% to 75.8% for the classifier), making large-model reasoning deployable in real time.
  • Compressing retrieval embeddings from 768 to 128 dimensions costs little recall, so the index can scale to billions of clips with lower storage cost.
  • Online A/B results attribute a 6–8% reduction in user views of unwanted livestreams to the combined system, with the reference path contributing a 2–4% additional reduction beyond classification.
  • Thresholding at P90 precision, human review at lower precision bands, and an appeal mechanism limit over-moderation.

Where Pith is reading between the lines

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

  • If the similarity path is validated, the same retrieval-plus-rerank design could generalize to other evolving-content domains (fraud, spam, dangerous behaviors) where policy shifts faster than training labels.
  • The paper does not quantify how detection degrades as the reference index ages; a testable extension is to replace the index with rolling windows of recent violations—if recall drops sharply, the system depends on a fresh reference set rather than on true generalization.
  • The reported 22% exclusive-coverage figure is not broken down by violation type; an external audit could check whether similarity-path detections are concentrated in a few categories, which would change how one interprets the 'novel violation' claim.
  • Since A/B metrics measure user views rather than precision/recall, the business effect may combine better detection with milder enforcement; separating the two would require instrumenting per-action outcomes.

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

5 major / 5 minor

Summary. The paper describes a production hybrid content-moderation system for livestreams, combining (1) a supervised multiclass classification pipeline for known violation categories and (2) a reference-based similarity matching pipeline that retrieves near neighbors from a curated index of historical violations and re-ranks them with a lightweight multimodal model. Both pipelines are trained with knowledge distillation from a fine-tuned MLLM (LLaVA-One-Vision) to keep inference light enough for 20-second clip processing. The authors report offline component metrics (classification AP 75.84%, re-ranking AP 74.82%, retrieval Recall@Top-100 98.99%), a runtime latency table (0.4 s vs 4 s per clip), and online A/B results claimed to show a 6–8% reduction in user views of unwanted livestreams, with reference matching said to add roughly 22% coverage beyond classification.

Significance. If the central claims were fully supported, this would be a valuable systems contribution: the dual-path architecture is a sensible engineering response to the live-stream moderation problem, the latency and deployment details are informative, and the distillation ablations (Tables 4, 5, 7) provide useful evidence that compact students can approach MLLM teachers. The paper also ships concrete algorithmic details (MoCo+CLIP, MRL, HNSW, aggregation) and a large-scale experimental footprint. However, the headline quantitative claims are not currently supported by the paper's own tables: the abstract's '76% recall at 80% precision' is contradicted by the reported re-ranking numbers, the '22% additional coverage' figure is never defined or measured, and the online A/B section contains internal inconsistencies. These issues bear directly on the paper's central contribution, so they must be resolved before the claims can be accepted.

major comments (5)
  1. [Abstract; §4.5, Table 3; §4.6.3, Table 7; Table 10] The abstract states the similarity pipeline 'achieves 76% recall at 80% precision,' but no table reports this value. §4.5/Table 3 gives the multimodal re-ranking model a Recall@P80 of 66.13%; Table 7 gives the distilled Small+KD model 66.13% and the full MLLM 72.54%; Table 10's best R@P80 is 70.55%. There is no configuration anywhere in the paper with R@P80=76%. The abstract's headline performance claim must either be corrected to match the reported experiments or accompanied by the specific operating point/experiment that produces 76%.
  2. [§1; §4 (all); §4.7] Section 1 asserts that 'reference matching contributes approximately 22% additional coverage beyond the classification branch.' No definition of 'coverage' is given, and no experiment or table in Section 4 measures the union of the two branches or the marginal contribution of reference matching to overall detection coverage. The online backtest in §4.7 reports reductions in user views (4% + 2–4%), which is a different metric. Without an end-to-end measurement of the hybrid system's coverage relative to classification alone, the complementarity claim that motivates the entire architecture is unsupported.
  3. [§4.7 Online Performance] The A/B statistics are internally inconsistent as written. The preset-detection experiment reports 'Metric: 1.2% decrease in violation-related user views' but then gives a 95% confidence interval of [−0.1476%, −0.087%], which excludes −1.2% and instead brackets about −0.12%. This is likely a transcription or unit error, but as written it is impossible to interpret. In addition, the component experiments (1.2% and 0.6%) are not reconciled with the long-term 'A/B backtest' numbers (4% + 2–4%) or the abstract's 6–8%, and no common baseline or metric definition is provided. The online impact claim needs a single consistent, labeled set of metrics and confidence statements.
  4. [Table 10] Table 10 is captioned as 'Score aggregation of the reference matching branch,' but its 'No aggr.' row exactly reproduces the preset violation detection results in Table 1 (AP 75.84%, F1 73.61%, R@P70 75.63%, R@P75 71.40%, R@P80 66.70%, R@P85 62.93%, R@P90 47.37%). This suggests the row is either mislabeled or copied from the wrong model. Consequently, the claimed improvement from aggregation (AP 75.84 → 78.49, R@P80 66.70 → 70.55) cannot be interpreted. The table must be corrected, and the no-aggregation baseline for the reference branch must be reported separately.
  5. [§3.3; §1] The reference-matching pipeline is justified as detecting 'novel, adversarial, or edge-case violations' by comparing incoming clips to a curated database of historical violations. This design only works if the reference index is representative of the emerging violation distribution or is continuously refreshed; the paper gives no data on index coverage, freshness, decay, or the novelty distribution of detected violations. The '22% additional coverage' figure, if supplied, should be analyzed by how many of those detections are genuinely novel versus near-duplicates of existing indexed violations. As it stands, the claimed ability to handle emerging violations is an untested assumption, not an empirical result.
minor comments (5)
  1. [Throughout] Several typos and formatting issues: 'complimentary' (§1), 'popluar' (§2.1), 'Figrue' in the captions of Figure 4 and Figure 6, 'constrastive' (§4.6.2), 'divergence Loss' in Figure 6. These should be corrected.
  2. [References [18]/[19]] References [18] and [19] both point to the same LLaVA-OneVision paper (arXiv:2408.03326), but are cited differently in the text ([18] in the caption of Figure 6, [19] elsewhere). Please unify.
  3. [§4.2; Table 5; Table 8] Section 4.2 says training used NVIDIA H100 GPUs, but Table 5 reports throughput in QPS on an A10 GPU. Clarify whether latency/throughput numbers are from A10 or H100, and whether Table 8's deployment latency is on the same hardware or production servers.
  4. [§4.7; Table 10; Algorithm 1] The main text says the deployed aggregation is 'simple sliding-window aggregation with mean pooling' (Table 10), while Appendix A.1 presents Algorithm 1 (Clip-Match Aggregation with temporal tolerance and score threshold). The relationship between the two is unclear: is Algorithm 1 an alternative/earlier method, and does Table 10 use Algorithm 1 or mean-pooling over windows? Please state explicitly.
  5. [Table 2] The 'Recall one' and 'Recall all' metrics in Table 2 are used without formal definitions in the main text. Please define them (e.g., fraction of queries with at least one relevant item in top-k, versus fraction of all relevant items retrieved).

Circularity Check

0 steps flagged

No significant circularity: an empirical systems paper evaluated against human labels and online view-rate metrics; minor self-citation is not load-bearing.

full rationale

This paper does not present a formal derivation, so the circularity review reduces to checking whether any headline number is a fitted input renamed as a prediction or justified only by a self-citation. I find neither. The component metrics (Tables 1, 2, 3, 4, 6, 7) are measured against human-annotated clip/pair labels; the distillation claims compare Small+KD with non-distilled Small and the MLLM teacher on the same human-labeled benchmark, so the teacher outputs are training signals, not the evaluation target (e.g., Table 4: Small+KD AP=75.84% vs. Small v+a AP=71.05%, evaluated independently). The online A/B results are live view-rate measurements, not quantities derived from the offline fitted scores. The only author-overlap citation with any methodological role is Info-Coevolution [28] for the active-learning sampler; active learning is peripheral to the paper's central dual-path claim and is not invoked as a uniqueness theorem or as the source of the headline reductions, so it is at most a minor self-citation. The '22% additional coverage' figure and the abstract's '76% recall at 80% precision' are not backed by a matching table entry and the table values conflict (Tables 3 and 7 give R@P80=66.13%; Table 10 maxes at 70.55%); these are evidentiary/consistency problems, not cases where a conclusion reduces by construction to its inputs. No equation in the paper defines a target metric in terms of itself, and no fitted parameter is relabeled as a prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central empirical claims rest on production data and operational choices that are not externally observable. The listed free parameters are thresholds and hyperparameters that control the reported precision-recall trade-offs; the axioms are the data and measurement assumptions behind the A/B evaluation. No new scientific entities are introduced.

free parameters (5)
  • Aggregation score threshold τ = not disclosed
    Algorithm 1 uses τ to decide whether a query-reference clip pair is a valid match; reported aggregation gains depend on this threshold.
  • Temporal tolerance ε = not disclosed
    Algorithm 1 uses ε as the temporal alignment window for clip-match aggregation; it controls the precision/recall trade-off of session-level matches.
  • Deployment precision operating point = P90 automated; P70–P85 human review
    Model thresholds are tuned at P90 precision for automated enforcement, with P70–P85 sent to human moderators; this directly determines the reported precision-recall figures.
  • MoCo/CLIP loss hyperparameters (α, β, λ, τ) = not disclosed
    The retrieval training objective in §3.3 depends on weighting hyperparameters and a similarity margin/temperature; these affect embedding quality and recall.
  • Clip segmentation length = 20 seconds
    Streams are split into fixed 20-second clips; Table 9 shows sensitivity to clip length, so this design choice is load-bearing for the aggregation results.
axioms (5)
  • domain assumption 20-second clips contain enough signal to detect violations
    §3.1 segments every livestream into 20-second query clips; if violating content requires longer context, detection and reranking will miss it.
  • domain assumption Human-annotated in-house labels are reliable ground truth
    §4.1 relies on 2M human-annotated clips and 0.5M annotated pairs; no inter-annotator agreement or label-quality analysis is reported.
  • domain assumption Reference index of historical violations is representative of future novel violations
    §3.3 assumes that comparing new clips to 'a curated database of known violations' generalizes to 'novel, adversarial, or edge-case violations'; coverage of unseen violation types is not measured.
  • domain assumption ASR transcription captures the audio content relevant to violations
    The re-ranking model uses ASR text as a key modality; violations carried in non-speech audio or in languages/accents poorly handled by ASR would be invisible to this branch.
  • domain assumption The A/B metric 'views of unwanted livestreams' isolates moderation quality
    §4.7 treats a reduction in this metric as evidence of better moderation, but it can be confounded by recommendation changes, user behavior shifts, and enforcement-side effects.

pith-pipeline@v1.3.0-alltime-deepseek · 15968 in / 13241 out tokens · 117729 ms · 2026-08-03T18:44:26.882330+00:00 · methodology

0 comments
read the original abstract

Content moderation remains a critical yet challenging task for large-scale user-generated video platforms, especially in livestreaming environments where moderation must be timely, multimodal, and robust to evolving forms of unwanted content. We present a hybrid moderation framework deployed at production scale that combines supervised classification for known violations with reference-based similarity matching for novel or subtle cases. This hybrid design enables robust detection of both explicit violations and novel edge cases that evade traditional classifiers. Multimodal inputs (text, audio, visual) are processed through both pipelines, with a multimodal large language model (MLLM) distilling knowledge into each to boost accuracy while keeping inference lightweight. In production, the classification pipeline achieves 67% recall at 80% precision, and the similarity pipeline achieves 76% recall at 80% precision. Large-scale A/B tests show a 6-8% reduction in user views of unwanted livestreams}. These results demonstrate a scalable and adaptable approach to multimodal content governance, capable of addressing both explicit violations and emerging adversarial behaviors.

Figures

Figures reproduced from arXiv: 2512.03553 by Danhui Guan, David Yuchen Wang, Hailun Xu, Hiok Hian Ong, Kanchan Sarkar, Sanjay Saha, Wei Chee Yew, Xiaotian Fan, Zhenheng Yang.

Figure 1
Figure 1. Figure 1: Overview of our proposed hybrid moderation framework. Incoming live-stream clips are segmented into 20-second [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the supervised classification [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training pipeline for the video-clip retrieval fea [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Small Re-ranking Model: the small/lightweight mul [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: We observe that raising precision from 80% to 90% requires a steep threshold increase, as ’violation’ and ’non￾violation’ scores heavily overlap in this range, indicating most hard cases lie here. And, both precision and recall re￾main stable around 75−80%, with minor fluctuations, demon￾strating model stability. A.4 Knowledge Distillation of Re-ranking model Llava One Vision MLP SigLip Query Frames Projec… view at source ↗
Figure 6
Figure 6. Figure 6: Knowledge Distillation of Re-ranking model: the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MLT-Dedup: Efficient Large-Scale Online Video Deduplication via Multi-Level Representations and Spatial-Temporal Matching

    cs.CV 2026-06 unverdicted novelty 4.0

    MLT-Dedup achieves 91% reduction in online video repetition rates at 90% precision and 5x indexing capacity using multi-level representations and differential feature-enhanced similarity on a real-world platform.

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Nouar AlDahoul, Myles Joshua Toledo Tan, Harishwar Reddy Kasireddy, and Yasir Zaki. 2024. Advancing Content Moderation: Evaluating Large Lan- guage Models for Detecting Sensitive Content Across Text, Images, and Videos. arXiv:2411.17123 [cs.CV] https://arxiv.org/abs/2411.17123

    Pith/arXiv arXiv 2024

  2. [2]

    Tommaso Caselli, Valerio Basile, Jelena Mitrović, and Michael Granitzer. 2021. HateBERT: Retraining BERT for Abusive Language Detection in English. arXiv:2010.12472 [cs.CL] https://arxiv.org/abs/2010.12472

    Pith/arXiv arXiv 2021

  3. [3]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning (ICML). PMLR, 1597–1607

    2020

  4. [4]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2024. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv:2312.14238 [cs.CV] https://arxiv.org/abs/2312.14238

    Pith/arXiv arXiv 2024

  5. [5]

    Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. Qwen-Audio: Advancing Univer- sal Audio Understanding via Unified Large-Scale Audio-Language Models. arXiv:2311.07919 [eess.AS] https://arxiv.org/abs/2311.07919

    Pith/arXiv arXiv 2023

  6. [6]

    Pedro V. A. de Freitas, Paulo R. C. Mendes, Gabriel N. P. dos Santos, Anto- nio José G. Busson, Álan Livio Guedes, Sérgio Colcher, and Ruy Luiz Milidiú

  7. [7]

    Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. 2022. An Empirical Study of Training End -to-End Vision-and-Language Transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 18166–18176

    2022

  8. [8]

    Nathan Matias

    Alexandra Gonzalez and J. Nathan Matias. 2025. Measuring the Mental Health of Content Reviewers, a Systematic Review. arXiv:2502.00244 [cs.CY] https: //arxiv.org/abs/2502.00244

    Pith/arXiv arXiv 2025

  9. [9]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum Contrast for Unsupervised Visual Representation Learning. InCVPR

    2020

  10. [10]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 770–778. doi:10.1109/CVPR. 2016.90

    doi:10.1109/cvpr 2016

  11. [11]

    Sifeng He, Yue He, Minlong Lu, Chen Jiang, Xudong Yang, Feng Qian, Xi- aobo Zhang, Lei Yang, and Jiandong Zhang. 2022. TransVCL: Attention- enhanced Video Copy Localization Network with Flexible Supervision. arXiv:2211.13090 [cs.CV] https://arxiv.org/abs/2211.13090

    Pith/arXiv arXiv 2022

  12. [12]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531(2015). doi:10.48550/arXiv. 1503.02531 Originally published as a NIPS 2014 Deep Learning Workshop paper

    Pith/arXiv arXiv doi:10.48550/arxiv.1503.02531 2015

  13. [13]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations (ICLR). Preprint available on OpenReview

    2022

  14. [14]

    Chen Jiang, Kaiming Huang, Sifeng He, Xudong Yang, Wei Zhang, Xiaobo Zhang, Yuan Cheng, Lei Yang, Qing Wang, Furong Xu, Tan Pan, and Wei Chu. 2021. Learning Segment Similarity and Alignment in Large-Scale Content Based Video Retrieval. InProceedings of the 29th ACM International Conference on Multimedia (MM ’21). ACM, 1618–1626. doi:10.1145/3474085.3475301

    arXiv 2021

  15. [15]

    Rohan Kshirsagar, Tyus Cukuvac, Kathleen McKeown, and Susan McGre- gor. 2018. Predictive Embeddings for Hate Speech Detection on Twitter. arXiv:1809.10644 [cs.CL] https://arxiv.org/abs/1809.10644

    Pith/arXiv arXiv 2018

  16. [16]

    Deepak Kumar, Yousef AbuHashem, and Zakir Durumeric. 2024. Watch Your Language: Investigating Content Moderation with Large Language Models. arXiv:2309.14517 [cs.HC] https://arxiv.org/abs/2309.14517

    Pith/arXiv arXiv 2024

  17. [17]

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and Ali Farhadi. 2022. Matryoshka Representation Learning. In Proceedings of NeurIPS 2022

    2022

  18. [19]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2024. LLaVA- OneVision: Easy Visual Task Transfer. arXiv:2408.03326 [cs.CV] https://arxiv. org/abs/2408.03326

    Pith/arXiv arXiv 2024

  19. [20]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 10012–10022

    2021

  20. [21]

    Minlong Lu, Yichen Lu, Siwei Nie, Xudong Yang, and Xiaobo Zhang. 2025. Self- supervised Video Copy Localization with Regional Token Representation. In Computer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga KDD ’26, August 09–13, 2026, Jeju Island, Republic of Korea Wei Chee Yew et al. Russakovsky, Torsten Sattler, and Gül Varol (Eds.)...

    2025

  21. [22]

    Xingyu Lu, Tianke Zhang, Chang Meng, Xiaobei Wang, Jinpeng Wang, YiFan Zhang, Shisong Tang, Changyi Liu, Haojie Ding, Kaiyu Jiang, Kaiyu Tang, Bin Wen, Hai-Tao Zheng, Fan Yang, Tingtingx Gao, Di Zhang, and Kun Gai. 2025. VLM as Policy: Common-Law Content Moderation Framework for Short Video Platform. arXiv:2504.14904 [cs.SI] https://arxiv.org/abs/2504.14904

    Pith/arXiv arXiv 2025

  22. [23]

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, Changqing Qiu, Jiaqi Zhang, Xu Zhang, Zhiheng Yan, Jingming Zhang, Simin Zhang, Mingxing Wen, Zhaojie Liu, Kun Gai, and Guorui Zhou. 2024. QARM: Quantitative Alignment Multi-Modal Recommendation at Kuaishou. arXiv:2411.11739 [cs.IR] htt...

    Pith/arXiv arXiv 2024

  23. [24]

    Henrietta Lyons, Eduardo Velloso, and Tim Miller. 2021. Conceptualising Contestability: Perspectives on Contesting Algorithmic Decisions.Proceed- ings of the ACM on Human-Computer Interaction5, CSCW1 (April 2021), 1–25. doi:10.1145/3449180

    doi:10.1145/3449180 2021

  24. [25]

    Malkov and Dmitry A

    Yury A. Malkov and Dmitry A. Yashunin. 2020. Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs.IEEE Transactions on Pattern Analysis and Machine Intelligence42, 4 (2020), 824–836

    2020

  25. [26]

    Taber, Andreas Damianou, and Mounia Lalmas

    Konstantina Palla, José Luis Redondo García, Claudia Hauff, Francesco Fabbri, Henrik Lindström, Daniel R. Taber, Andreas Damianou, and Mounia Lalmas

  26. [27]

    Tianyu Peng and Jiajun Zhang. 2024. Enhancing Knowledge Distillation of Large Language Models through Efficient Multi-Modal Distribution Alignment. InProceedings of COLING

    2024

  27. [28]

    Ziheng Qin, Hailun Xu, Wei Chee Yew, Qi Jia, Yang Luo, Kanchan Sarkar, Danhui Guan, Kai Wang, and Yang You. 2025. Info-Coevolution: An Efficient Framework for Data Model Coevolution. InProceedings of the 42nd International Conference on Machine Learning (ICML). Poster 45070; machine learning, coevolution of data and models

    2025

  28. [29]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InProceedings of the 38th International Conference on Machine Learning (ICML). PMLR, 8748–8763

    2021

  29. [30]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervi- sion. arXiv preprint arXiv:2212.04356. Includes Whisper-Tiny, Whisper -Base, Whisper-Small, etc

    Pith/arXiv arXiv 2022

  30. [31]

    2019.Behind the Screen: Content Moderation in the Shadows of Social Media

    Sarah Roberts. 2019.Behind the Screen: Content Moderation in the Shadows of Social Media. doi:10.2307/j.ctvhrcz0v

    work page doi:10.2307/j.ctvhrcz0v 2019

  31. [32]

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Lud- wig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. 2022. LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image -Tex...

    Pith/arXiv arXiv 2022

  32. [33]

    Statista Research Department. 2025. Live Streaming - Statistics & Facts. https: //www.statista.com/topics/8906/live-streaming/. Accessed: 2025-08-01

    2025

  33. [34]

    Siqi Sun, Zhe Gan, Yu Cheng, Shuohang Wang, and Jingjing Liu. 2020. Contrastive Distillation on Intermediate Representations for Language Model Compression. arXiv preprint arXiv:2009.14167(2020)

    Pith/arXiv arXiv 2020

  34. [35]

    Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530 [cs.CL] https://arxiv.org/abs/ 2403.05530

    Pith/arXiv arXiv 2024

  35. [36]

    OpenAI Team. 2024. GPT-4o System Card. arXiv:2410.21276 [cs.CL] https: //arxiv.org/abs/2410.21276

    Pith/arXiv arXiv 2024

  36. [37]

    TikTok Newsroom. 2024. Celebrating the Power of Community and Creativity on TikTok LIVE in 2024. https://newsroom.tiktok.com/en-us/celebrating-the- power-of-community-and-creativity-on-tiktok-live-in-2024. Accessed: 2025- 08-01

    2024

  37. [38]

    Xiaohan Xu, Ming Li, Chongyang Tao, and Tao Shen. 2024. A Survey on Knowl- edge Distillation of Large Language Models.arXiv preprint arXiv:2402.13116 (2024)

    Pith/arXiv arXiv 2024

  38. [39]

    Wenjun Zeng, Yuchi Liu, Ryan Mullins, Ludovic Peran, Joe Fernandez, Hamza Harkous, Karthik Narasimhan, Drew Proud, Piyush Kumar, Bhaktipriya Rad- harapu, Olivia Sturman, and Oscar Wahltinez. 2024. ShieldGemma: Gener- ative AI Content Moderation Based on Gemma. arXiv:2407.21772 [cs.CL] https://arxiv.org/abs/2407.21772

    Pith/arXiv arXiv 2024

  39. [40]

    Zirui Zhu, Hailun Xu, Yang Luo, Yong Liu, Kanchan Sarkar, Zhenheng Yang, and Yang You. 2025. FOCUS: Efficient Keyframe Selection for Long Video Under- standing.arXiv preprint arXiv:2510.27280(2025). A Appendix A.1 Clip-Match Aggregation Algorithm Post-processing step to improve re-ranking precision by verifying prolonged matches between a query livestream...

    arXiv 2025

  40. [2019]

    arXiv:1911.03974 [cs.MM] https://arxiv.org/abs/1911.03974

    A Multimodal CNN-based Tool to Censure Inappropriate Video Scenes. arXiv:1911.03974 [cs.MM] https://arxiv.org/abs/1911.03974

    Pith/arXiv arXiv 1911

  41. [2025]

    arXiv:2502.18695 [cs.CY] https://arxiv.org/abs/2502.18695

    Policy-as-Prompt: Rethinking Content Moderation in the Age of Large Language Models. arXiv:2502.18695 [cs.CY] https://arxiv.org/abs/2502.18695

    Pith/arXiv arXiv