Pith. sign in

REVIEW 4 major objections 5 minor 31 references

TANDEM converts long-video hate speech detection into a structured reasoning task that localizes hate segments and identifies their targets, reporting 0.73 F1 for target identification on HateMM.

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 10:04 UTC pith:REQFJAOW

load-bearing objection New structured-output task framing on HateMM/MHC, but the SCCR/tandem causal claims are untested—the missing ablations and a statistically flat headline target-F1 undermine the abstract. the 4 major comments →

arxiv 2601.11178 v3 pith:REQFJAOW submitted 2026-01-16 cs.AI cs.CLcs.MMcs.SI

TANDEM: Temporal-Aware Neural Detection for Multimodal Hate Speech

Girish A. Koushik , Helen Treharne , Diptesh Kanojia This is my paper
classification cs.AI cs.CLcs.MMcs.SI
keywords multimodal hate speech detectiontemporal localizationtarget identificationreinforcement learningvision-language modelaudio-language modellong-form videostructured prediction
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 is trying to establish that hate speech moderation for long-form social media video should be — and can be — a structured reasoning problem rather than a binary flag. TANDEM fine-tunes a vision-language model and an audio-language model in alternating reinforcement-learning phases, each conditioning the other through self-generated cross-modal context, and rewards the pair for correct classification, accurate hate-segment timestamps, and correct target-group identification. On the HateMM dataset the complete pipeline reaches 0.73 F1 for target identification, about 30% above the strongest zero-shot baseline, with temporal grounding that zero-shot models struggle to match. The authors' larger claim is that interpretable, structured alignment is achievable in long, noisy, multimodal content without dense frame-level supervision, which matters because moderators need evidence — where and at whom — rather than a single score.

Core claim

TANDEM's central claim is that audio-visual hate understanding can be reframed as structured reasoning: for each 30-second chunk of a long video, a vision-language model and an audio-language model each emit an XML-like prediction containing a classification (Hate/Non-hate or Normal/Offensive/Hateful), timestamps of hateful spans, target-group labels, and a short rationale. The models are trained in what the paper calls a tandem reinforcement learning strategy: during a phase, one model is updated while the other is frozen and supplies structured context produced by a zero-shot 'self-constrained context round' (SCCR); then the roles swap. A composite reward combines cross-entropy on classifi

What carries the argument

The load-bearing mechanism is the self-constrained context round (SCCR) combined with alternating policy updates. Before an RL phase, the frozen modality performs a zero-shot inference pass to generate structured context — audio context for the vision-language model, visual context for the audio-language model — and the trainable model is optimized while conditioning on that context, with roles swapped every 10 steps. The paper's reward function is the second central object: R = λ_c CE + λ_τ IoU + λ_z F1 + λ_f R_len + λ_f R_fmt, which ties policy updates directly to the structured outputs (classification, timestamps, targets, format). SCCR is meant to reduce reward sparsity and prevent modal

Load-bearing premise

The load-bearing premise is that the SCCR context and the alternating RL updates cause the reported gains; the ablations never remove SCCR from an RL run, and the SFT-only audio model already reaches 0.71±0.11 target F1, within noise of the 0.73 headline, so if supervision alone accounts for the improvement the tandem mechanism's role is unsupported.

What would settle it

Run the same SFT + GRPO/GSPO recipe on HateMM with SCCR removed (and with RL only, without SFT), keeping seeds identical; if target F1 stays at 0.73±0.11, SCCR contributes nothing. Conversely, a fully supervised segment-and-target model trained directly on HateMM's dense annotations that matches or exceeds 0.73 would undercut the claim that RL-based structured alignment is needed.

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

If this is right

  • Human-in-the-loop moderators can be shown the predicted hate segment and target group together, making verification faster than reviewing coarse flags.
  • Because the same pipeline handles binary, multi-class, and implicit-hate label schemas across three datasets, a single trained framework could serve multiple moderation tiers.
  • Temporal grounding improves most when RL is built on top of SFT and SCCR, so training-side stabilization — not just bigger base models — is what unlocks structured outputs.
  • The remaining gap between hateful and offensive (or explicit and implicit) classes is attributed to label ambiguity and imbalance rather than detection failure, pointing dataset builders toward cleaner label schemas.
  • Long-video processing via 30-second chunks and scene-change keyframes means the approach scales to hours of content without dense frame annotations.

Where Pith is reading between the lines

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

  • The '≈30% improvement over state-of-the-art' is measured against zero-shot and context-augmented baselines, not against a supervised segment-localization model trained on HateMM's own annotations; a true supervised baseline could narrow or reframe that lead.
  • Since the SFT-only audio condition already scores 0.71±0.11 target F1, the marginal gain from SCCR/tandem RL may be small; an ablation that removes SCCR (or RL) while keeping SFT would be needed to isolate their contribution.
  • The SFT data is filtered by Qwen3-Omni predictions, so part of the observed gain may come from the teacher model's quality rather than from the tandem training scheme itself.
  • The paper's own checklist (item 4a) states code will be released only upon acceptance, which currently prevents independent reproduction of the RL configuration; until then the exact training recipe is unverifiable.

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

4 major / 5 minor

Summary. TANDEM proposes a framework for structured multimodal hate-speech understanding in long videos, jointly producing classification, temporal segments, and target identities. The pipeline chunks videos into 30-second pieces, uses Qwen2.5-VL-7B and Qwen2-Audio-7B with LoRA, performs SFT on a curated 100-video subset, and then applies GRPO/GSPO reinforcement learning with a self-constrained cross-modal context round (SCCR). The authors evaluate on HateMM, MultiHateClip, and ImpliHateVid, reporting a target-identification F1 of 0.73 on HateMM and claiming a roughly 30% improvement over state-of-the-art, while also claiming precise temporal grounding and interpretable structured outputs.

Significance. If the central claims were fully supported, TANDEM would be a valuable contribution to interpretable multimodal moderation and to the use of RL for structured prediction in long-form audio-visual content. The paper has clear strengths: it tackles an important and under-addressed problem, evaluates across three datasets, reports error bars over three seeds, includes multiple zero-shot and context-augmented baselines, and provides qualitative examples. These are meaningful. However, the load-bearing novelty — the SCCR and alternating tandem RL mechanism — is not isolated in the ablations, and several headline statements in the abstract and discussion go beyond what the tables show. The significance of the work therefore depends on additional experiments and a careful revision of the claims.

major comments (4)
  1. [Table 3, §4.3] The paper's central mechanistic claim is that SCCR and the tandem RL strategy cause the performance gains, yet Table 3 contains no RL run without SCCR. The rows labeled "SCCR + GSPO" and "SCCR + GRPO" include SCCR by construction, and all RL rows condition on SCCR. The statement in §4.3 that "Models trained with SCCR consistently outperform their counterparts trained without it" is therefore unsupported by any comparison in the table. Please add ablations: SFT + RL without SCCR, RL-only without SCCR, and ideally a joint-update (non-alternating) RL control, so that the specific contributions of SCCR and of the alternating tandem update are separately identified.
  2. [Table 3, §4.2] The SFT-only audio row already achieves target Avg F1 0.71±0.11 and Exact Match 0.58±0.13, while the full SFT + SCCR + GRPO audio row achieves 0.73±0.11 and 0.59±0.13. These differences are well within the reported error bars, so the abstract's claim that the full method "significantly" improves target identification, and the text's statement that it "reliably achieves high target identification," are not supported by the reported statistics. Please provide significance tests or effect sizes, or state plainly that the additional RL/SCCR components do not yield a statistically distinguishable gain on this metric.
  3. [Abstract, §1] The abstract claims "a ≈30% improvement over state-of-the-art" for target identification on HateMM, but no comparison baseline for that number is identified. In Table 2, the strongest zero-shot target Avg F1 is Gemini-2.5-Flash at 0.42; TANDEM's 0.73 is about 74% higher, not 30%. If the 30% refers to a different baseline or a different metric, that baseline and computation must be given explicitly. As written, the headline improvement is unverifiable and appears inconsistent with the paper's own table.
  4. [Tables 2 and 3, §4.3, §5] The paper repeatedly claims precise temporal grounding and says TANDEM "narrows this gap" or improves temporal localization, but the best TANDEM Avg IoU on HateMM is 0.43 (SFT + SCCR + GRPO, VL), which is below the zero-shot Qwen3-Omni baseline of 0.53. On MultiHateClip, the best TANDEM Avg IoU (0.13) ties the Qwen3-Omni baseline rather than exceeding it. These results undermine the "maintaining precise temporal grounding" claim in the abstract and the narrative in §4.3 and §5 that TANDEM improves temporal grounding. Please either report a setting where TANDEM surpasses the zero-shot localization baseline or temper the claims to reflect that temporal localization remains below a strong zero-shot model.
minor comments (5)
  1. [Table 3, Cfg legend] The legend uses w=[λ_len, λ_fmt, λ_c, λ_τ, λ_z], but Eq. (2) uses λ without a subscript for summary length and formatting. Please align the notation between the equation and the table legend.
  2. [Table 3, row labels] The row labels "SCCR + GSPO" and "SCCR + GRPO" are ambiguous because the text in §4.2 describes these as RL-only runs without SFT, while the following rows are labeled "SFT + SCCR + ...". Please make the presence or absence of SFT explicit in every row label.
  3. [§3.3, Eq. (1)] In Eq. (1), the notation π_θ_a and π_θ_v is introduced but the conditioning variables for the audio policy are not fully written out. Consider adding a brief explanation of the notation and of how the frozen modality's output is passed to the other modality.
  4. [Abstract] The phrase "significantly outperforms" appears without any statistical significance testing reported. Either provide significance tests or replace "significantly" with "outperforms."
  5. [§4.3] The discussion states that "The strongest and most consistent results are obtained with SFT + SCCR + RL," but Table 3 shows substantial overlap between configurations. Please qualify this statement with the measured effect sizes or confidence intervals.

Circularity Check

0 steps flagged

No significant circularity: results are empirical test-set evaluations; SCCR/tandem attribution gaps are experimental-control issues, not definitional reductions.

full rationale

TANDEM's derivation chain is empirical rather than definitional. The structured-output results in Table 3 are obtained by training on train splits and evaluating on held-out test folds (§3.4: 'models process test videos in 30-second chunks'). Eq. (2) rewards CE/IoU/F1 on training targets, and Eq. (3) is a GRPO/GSPO update; these define the training objective, not the reported test outcome. The SCCR context in Eq. (1) is a conditioning variable, not a restatement of the evaluation metric. The only same-author citation (Koushik, Kanojia, and Treharne 2025) is an externally published classification baseline in Table 2 with no target-ID number; it is not used to define the target-ID or localization claims. The skeptic's concerns—SFT-only audio already reaching 0.71±0.11 Avg F1 vs 0.73±0.11 for SFT+SCCR+GRPO, no RL-without-SCCR control, best TANDEM localization below Qwen3-Omni zero-shot, and the unstated SOTA comparator for the '≈30% improvement'—are statistical attribution, control, and verifiability problems, not circular reductions. The paper itself acknowledges variance ('performance overlaps ... are often within the margin of error') and reward sensitivity in Limitations, which further indicates honest empirical reporting rather than a self-referential derivation. No step in the paper reduces a predicted quantity to a fitted input by construction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or conceptual entities; it is an engineering combination of existing models and RL algorithms. The free parameters are manual reward weights and chunking choices that directly affect the reported metrics.

free parameters (4)
  • Reward weights λ (config a) = [1,1,1,1,1] (length, format, class, IoU, target)
    Chosen manually per configuration; no sensitivity analysis beyond configs a,b,c.
  • Reward weights λ (config c) = [0.15,0.15,0.3,0.2,0.2]
    Ad hoc re-weighting that changes results; no stated selection criterion.
  • Chunk size = 30 seconds
    Fixed by hand; affects temporal IoU and cross-boundary handling.
  • SFT silver subset size = 100 videos
    Chosen by compute; candidate annotations filtered by Qwen3-Omni accuracy vs ground truth.
axioms (4)
  • domain assumption HateMM/MHC/IHV annotations (labels, timestamps, targets) are correct and complete
    All metrics and conclusions assume ground truth accuracy; no annotation-quality analysis is provided.
  • domain assumption 30-second chunking with scene-change keyframes preserves all salient hate content for IoU evaluation
    Chunk-boundary truncation and keyframe sampling could miss brief hate segments; §3.1.
  • domain assumption GRPO/GSPO with the given rewards produces stable convergence for 200 steps on two A100s
    No training curves or convergence checks are shown; §4.2 admits reward sensitivity.
  • domain assumption XML-formatted outputs from Qwen models are parseable and faithfully reflect reasoning
    Structured outputs require valid XML; no parser error counts are reported.

pith-pipeline@v1.3.0-alltime-deepseek · 16713 in / 16093 out tokens · 154281 ms · 2026-08-03T10:04:32.916374+00:00 · methodology

0 comments
read the original abstract

Social media platforms are increasingly dominated by long-form multimodal content, where harmful narratives are constructed through a complex interplay of audio, visual, and textual cues. While automated systems can flag hate speech with high accuracy, they often function as "black boxes" that fail to provide the granular, interpretable evidence, such as precise timestamps and target identities, required for effective human-in-the-loop moderation. In this work, we introduce TANDEM, a unified framework that transforms audio-visual hate detection from a binary classification task into a structured reasoning problem. Our approach employs a novel tandem reinforcement learning strategy where vision-language and audio-language models optimize each other through self-constrained cross-modal context, stabilizing reasoning over extended temporal sequences without requiring dense frame-level supervision. Experiments across three benchmark datasets demonstrate that TANDEM significantly outperforms zero-shot and context-augmented baselines, achieving 0.73 F1 in target identification on HateMM (a 30% improvement over state-of-the-art) while maintaining precise temporal grounding. We further observe that while binary detection is robust, differentiating between offensive and hateful content remains challenging in multi-class settings due to inherent label ambiguity and dataset imbalance. More broadly, our findings suggest that structured, interpretable alignment is achievable even in complex multimodal settings, offering a blueprint for the next generation of transparent and actionable online safety moderation tools.

Figures

Figures reproduced from arXiv: 2601.11178 by Diptesh Kanojia, Girish A. Koushik, Helen Treharne.

Figure 1
Figure 1. Figure 1: Sample Video Frames from HateMM Dataset (Das [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The TANDEM Architecture. The system employs a tandem reinforcement learning strategy where vision and audio models are iteratively updated while conditioning on each other’s context. The Long Video Processor handles segmentation, while the Multi-Task Reward Function computes a composite score based on summary length, output formatting, classification accuracy, segment localization, and target identificatio… 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

31 extracted references · 8 linked inside Pith

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; et al. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923

    Pith/arXiv arXiv 2025

  4. [4]

    Chu, Y.; Xu, J.; Yang, Q.; Wei, H.; Wei, X.; Guo, Z.; Leng, Y.; Lv, Y.; He, J.; Lin, J.; et al. 2024. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759

    Pith/arXiv arXiv 2024

  5. [5]

    Cinelli, M.; De Francisci Morales, G.; Galeazzi, A.; Quattrociocchi, W.; and Starnini, M. 2021. The echo chamber effect on social media. Proceedings of the national academy of sciences, 118(9): e2023301118

    2021

  6. [6]

    Comanici, G.; Bieber, E.; Schaekermann, M.; Pasupat, I.; Sachdeva, N.; Dhillon, I.; Blistein, M.; Ram, O.; Zhang, D.; Rosen, E.; et al. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261

    Pith/arXiv arXiv 2025

  7. [7]

    Das, M.; Raj, R.; Saha, P.; Mathew, B.; Gupta, M.; and Mukherjee, A. 2023. Hatemm: A multi-modal dataset for hate video classification. In Proceedings of the International AAAI Conference on Web and Social Media, volume 17, 1014--1023

    2023

  8. [8]

    Erickson, J.; and Yan, B. 2025. Content Moderation and Hate Speech on Alternative Platforms: A Case Study of BitChute. Proceedings of the ACM on Human-Computer Interaction, 9(2): 1--18

    2025

  9. [9]

    C.; and Ghanem, B

    Escorcia, V.; Caba Heilbron, F.; Niebles, J. C.; and Ghanem, B. 2016. Daps: Deep action proposals for action understanding. In European conference on computer vision, 768--784. Springer

    2016

  10. [10]

    Guo, Y.; Liu, J.; Li, M.; Cheng, D.; Tang, X.; Sui, D.; Liu, Q.; Chen, X.; and Zhao, K. 2025. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 3302--3310

    2025

  11. [11]

    Guo, Y.; Liu, J.; Li, M.; Liu, Q.; Chen, X.; and Tang, X. 2024. Trace: Temporal grounding video llm via causal event modeling. arXiv preprint arXiv:2410.05643

    Pith/arXiv arXiv 2024

  12. [12]

    S.; Sharma, S.; Cao, R.; Nandi, P.; Nakov, P.; Chakraborty, T.; and Lee, R

    Hee, M. S.; Sharma, S.; Cao, R.; Nandi, P.; Nakov, P.; Chakraborty, T.; and Lee, R. K.-W. 2024. Recent advances in online hate speech moderation: Multimodality and the role of large models. Findings of the Association for Computational Linguistics: EMNLP 2024, 4407--4419

    2024

  13. [13]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3

    2022

  14. [14]

    Kentmen-Cin, C. 2025. Hate Speech on Social Media: A Systemic Narrative Review of Political Science Contributions. Social Sciences, 14(10): 610

    2025

  15. [15]

    A.; Kanojia, D.; and Treharne, H

    Koushik, G. A.; Kanojia, D.; and Treharne, H. 2025. Towards a Robust Framework for Multimodal Hate Detection: A Study on Video vs. Image-based Content. In Companion Proceedings of the ACM on Web Conference 2025, 2014--2023

    2025

  16. [16]

    Lambert, N.; Morrison, J.; Pyatkin, V.; Huang, S.; Ivison, H.; Brahman, F.; Miranda, L. J. V.; Liu, A.; Dziri, N.; Lyu, S.; et al. 2024. Tulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124

    Pith/arXiv arXiv 2024

  17. [17]

    Li, Z.; Yu, W.; Huang, C.; Liu, R.; Liang, Z.; Liu, F.; Che, J.; Yu, D.; Boyd-Graber, J.; Mi, H.; et al. 2025 a . Self-rewarding vision-language model via reasoning decomposition. arXiv preprint arXiv:2508.19652

    Pith/arXiv arXiv 2025

  18. [18]

    Li, Z.; Zhang, X.; Guo, Y.; Bennamoun, M.; Boussaid, F.; Dwivedi, G.; Gong, L.; and Ke, Q. 2025 b . Watch and Listen: Understanding Audio-Visual-Speech Moments with Multimodal LLM. arXiv preprint arXiv:2505.18110

    arXiv 2025

  19. [19]

    OpenAI, T. 2023. GPT-4V(ision) System Card. Technical report

    2023

  20. [20]

    Rehman, M. Z. U.; Bhatnagar, A.; Kabde, O.; Bansal, S.; and Kumar, N. 2025. ImpliHateVid: A benchmark dataset and two-stage contrastive learning framework for implicit hate speech detection in videos. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 17209--17221

    2025

  21. [21]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    Pith/arXiv arXiv 2024

  22. [22]

    Shou, Z.; Wang, D.; and Chang, S.-F. 2016. Temporal action localization in untrimmed videos via multi-stage cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1049--1058

    2016

  23. [23]

    Takada, T.; Suzuki, Y.; Takushima, H.; Tanoue, H.; Sato, H.; Kumar, A.; Nishihara, H.; Hori, T.; and Ueki, K. 2024. Direct Metric Optimization for Image Captioning through Reward-Weighted Augmented Data Utilization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8333--8346

    2024

  24. [24]

    Wang, H.; Wang, Z.; and Lee, R. K.-W. 2025. HateClipSeg: A Segment-Level Annotated Dataset for Fine-Grained Hate Video Detection. In Proceedings of the 33rd ACM International Conference on Multimedia, 13304--13310

    2025

  25. [25]

    R.; Naseem, U.; and Lee, R

    Wang, H.; Yang, T. R.; Naseem, U.; and Lee, R. K.-W. 2024. Multihateclip: A multilingual benchmark dataset for hateful video detection on youtube and bilibili. In Proceedings of the 32nd ACM International Conference on Multimedia, 7493--7502

    2024

  26. [26]

    Xing, L.; Dong, X.; Zang, Y.; Cao, Y.; Liang, J.; Huang, Q.; Wang, J.; Wu, F.; and Lin, D. 2025. Caprl: Stimulating dense image caption capabilities via reinforcement learning. arXiv preprint arXiv:2509.22647

    arXiv 2025

  27. [27]

    Xu, J.; Guo, Z.; Hu, H.; Chu, Y.; Wang, X.; He, J.; Wang, Y.; Shi, X.; He, T.; Zhu, X.; Lv, Y.; Wang, Y.; Guo, D.; Wang, H.; Ma, L.; Zhang, P.; Zhang, X.; Hao, H.; Guo, Z.; Yang, B.; Zhang, B.; Ma, Z.; Wei, X.; Bai, S.; Chen, K.; Liu, X.; Wang, P.; Yang, M.; Liu, D.; Ren, X.; Zheng, B.; Men, R.; Zhou, F.; Yu, B.; Yang, J.; Yu, L.; Zhou, J.; and Lin, J. 20...

    2025

  28. [28]

    Yue, J.; Yang, S.; Chen, T.; Jiao, J.; and Fu, Z. 2025. Multimodal Hate Detection Using Dual-Stream Graph Neural Networks. arXiv preprint arXiv:2509.13515

    arXiv 2025

  29. [29]

    Zannettou, S.; Caulfield, T.; Blackburn, J.; De Cristofaro, E.; Sirivianos, M.; Stringhini, G.; and Suarez-Tangil, G. 2018. On the origins of memes by means of fringe web communities. In Proceedings of the internet measurement conference 2018, 188--202

    2018

  30. [30]

    Zhao, Y.; Xiong, Y.; Wang, L.; Wu, Z.; Tang, X.; and Lin, D. 2017. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision, 2914--2923

    2017

  31. [31]

    Zheng, C.; Liu, S.; Li, M.; Chen, X.-H.; Yu, B.; Gao, C.; Dang, K.; Liu, Y.; Men, R.; Yang, A.; et al. 2025. Group sequence policy optimization. arXiv preprint arXiv:2507.18071

    Pith/arXiv arXiv 2025