Pith. sign in

REVIEW 3 major objections 6 minor 77 references

Reading Between the Frames: Interpreting Implicit and Non-literal Meaning in Social Media Videos

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper introduces DrivelHub+, a 1,000-video benchmark of seemingly nonsensical social-media clips, and presents evidence that current video-language models can describe what these videos show but usually fail to explain what they mean.

desk verdict A useful new benchmark for implicit pragmatic video meaning, undermined by unvalidated gold explanations and an uncalibrated LLM judge—fixable before the numbers can be trusted. read the letter →

arxiv 2608.04939 v1 pith:NS647DFB submitted 2026-08-05 cs.CL

Yang Wang , Yanan Ma , Yiqi Liu , Zi Yan Chang , Chi-Li Chen , Chia-Yi Hsiao , Tyler Loakman , Aline Villavicencio
show 2 more authors
Chenghao Xiao Chenghua Lin
This is my paper
classification cs.CL
keywords multimodaldrivelologyimplicitmeaningunderstandingvideo-languagemodelspragmaticinterpretationsocialmediavideosreasoning-as-retrievalbenchmarkevaluationnon-literal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces DrivelHub+, a benchmark of 1,000 short-form social-media videos that look trivial or absurd on the surface but are deliberately structured to carry an implicit humorous, satirical, or socially pointed meaning, each paired with a human-written explanation of that hidden meaning. The benchmark asks whether video-language models can move from describing what is shown to inferring what is meant, and it evaluates them in two complementary ways: free-form explanation scored against the human annotations, and bidirectional video-to-text and text-to-video retrieval that tests whether videos and their interpretations sit close together in representation space. The paper reports that the strongest explanation model reaches about 9.5 of 12 on a rubric-based judge score and the strongest retrieval model about 75 percent Recall@1, while most systems fall far lower, and takes this as evidence of a persistent gap between surface-level recognition and pragmatic interpretation.

What carries the argument

The central object is the DrivelHub+ benchmark: 1,000 social-media videos, each carrying a human-written implicit-narrative explanation and metadata recording which modality — on-screen text, audio, visual content, or a combination — supplies the evidence for the interpretation. The argument is carried by two evaluation mechanisms. The explanation setting uses a video-LLM-as-a-judge protocol whose 12-point rubric rewards recovering the core intent, the rhetorical mechanism, the social or affective implication, and the multimodal grounding of the video, while penalising hallucination, literal-only description, and vagueness. The representation setting adapts reasoning-as-retrieval into bidirectional video-text retrieval, checking whether a video and its written interpretation are neighbours in embedding space rather than matched only by surface lexical or visual similarity.

What would settle it

Score a random subset of about 100 model-generated explanations with the same rubric using multiple human raters and compare their labels and totals with the judge model's. If human-judge agreement is low (Cohen's kappa below roughly 0.5) or if the human-ranked order of systems differs from the judge-ranked order, the central claim of a persistent perception-pragmatics gap would need revision, because the gap is measured entirely through this single judge.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that current video-language models can often recognise the visible content of drivelological videos — the people, actions, captions, and speech — while still failing to recover the implicit narrative that makes the clip meaningful, and that this failure persists across both open-ended explanation and representation-level retrieval. The best explanation model scores about 9.5 of 12 on the judge rubric, and the best retrieval model reaches about 75 percent Recall@1 in the text-to-video direction, but many evaluated systems, including video LLMs adapted for retrieval, perform far below these levels, with some near chance. The paper also claims that generation and representation are distinct abilities: retrieval-native embedding models vastly outperform generative video LLMs adapted for the same retrieval task, so articulating an interpretation and aligning a video with its interpretation in embedding space are not the same capability and should be evaluated separately.

Load-bearing premise

The explanation results rest on the assumption that one judge model, Qwen3.6-35B-A3B, gives valid alignment labels and rubric scores for every system with no calibration against human ratings and no agreement study; since other models from the same family are among the systems being judged, any family-specific bias would distort the reported gap.

Editorial extensions

If this is right

  • Current video-language models are far from saturated on implicit pragmatic meaning, so benchmark results that track surface description alone overstate what models actually understand about social-media video.
  • Improving a model's ability to explain a joke will not automatically improve its ability to align videos with meanings in representation space, since the two capabilities diverge sharply in the paper's results; both need separate training and evaluation.
  • Vision is the load-bearing input for these interpretations, including for on-screen text read visually, while audio plays a selective role and can even distract, so future models should prioritise visual and OCR-like evidence for pragmatic inference.
  • Test-time reasoning helps only when it stays grounded in the decisive visual or textual cue; models that reason at length but miss the semantic trigger, such as a canvas switching from portrait to landscape orientation, still fail.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The benchmark assumes each video has one recoverable intended meaning, yet the 85 percent agreement on inclusion decisions suggests the phenomenon's boundary is soft; an interesting extension would be to measure how often human viewers disagree on what a clip means, not just on what it shows.
  • Because the paper attributes LCO-Omni's retrieval gains to contrastive refinement of a strong generative backbone, a testable prediction is that applying the same contrastive training to the stronger Qwen3.5 or Qwen3.6 generation models would push retrieval Recall@1 well beyond the current roughly 75 percent.
  • The judge-based absolute scores would be worth re-checking with human preference data or a second judge model; the relative ordering of models is the more defensible part of the evidence, while the precise size of the reported gap may shift.
  • The benchmark's modality and language metadata make it a ready-made diagnostic for cultural transfer: re-running the protocol on the non-English speech and caption subsets would quantify how much of pragmatic understanding depends on culturally shared knowledge.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces DrivelHub+, a benchmark of 1,000 short-form social-media videos paired with human-written implicit-narrative explanations, and evaluates current video-language models on two tasks: free-form explanation of the implicit pragmatic meaning, and bidirectional video-text retrieval. The authors report that state-of-the-art open models reach only about 76% alignment and 9.5/12 judge score on explanation, and about 75% R@1 on retrieval, with generative video LLMs lagging far behind embedding-native retrieval models. They also present modality-ablation results showing vision is more critical than audio, and case studies illustrating frame-level reasoning failures. The central conclusion is that current models can describe surface content but often fail to infer what is meant.

Significance. If the evaluation is valid, DrivelHub+ is a useful diagnostic benchmark for an underexplored capability: pragmatic, culturally situated multimodal interpretation. The two-level protocol—explanation plus reasoning-as-retrieval—is a strong design choice, because retrieval provides an objective, judge-free signal that complements LLM-based scoring. The paper also contributes a broad model sweep, a clear modality-annotation scheme, and several concrete qualitative analyses. However, the reliability of the gold explanations and the validity of the single LLM judge are not yet established, and both are load-bearing for the headline numbers. With those validated, the benchmark and its findings would be a solid contribution; in the current form, the quantitative claims should be treated as provisional.

major comments (3)
  1. [Section 3.3 / Appendix D.1] The gold implicit-meaning explanations are written by five annotators who are research-team members or close collaborators, and the only third-party reliability check reports 85% agreement on inclusion decisions for 100 videos, not on the content of the explanations themselves. Because the central claim is that models fail to infer 'what is meant,' and the Limitations section acknowledges that the phenomenon is 'inherently interpretive,' the target annotations need demonstrated inter-annotator agreement on the meaning content. Without such evidence, a model that produces a valid alternative interpretation is scored as unaligned, so the reported gap between surface description and pragmatic inference may partly reflect disagreement with one annotation team's subjective readings. The authors should have independent annotators write explanations for a sample, measure agreement (e.g., semantic alignment between human explanations), and report the result; releasing multiple gold explanations per video would further strengthen the benchmark.
  2. [Section 4.2 / Appendix E.2] All explanation scores in Tables 2 and 10 are produced by a single judge, Qwen3.6-35B-A3B, which is also from the same model family as several evaluated systems. The paper provides no human calibration or agreement study showing that this judge's alignment labels and rubric scores correspond to human judgments. Since the judge receives the human annotation and the model explanation and is asked to judge semantic alignment, any systematic judge bias—such as favoring or penalizing particular stylistic choices or reasoning patterns—directly distorts the headline alignment rate and total score. The authors should validate the judge against human ratings on a sample, report inter-annotator agreement with humans, and ideally use multiple judges or a human-verified subset to confirm the main comparisons.
  3. [Section 4.3 / Appendix E.3] The retrieval relevance judgments are defined by treating the original video-narrative pair as always relevant and adding 'additional relevant pairs' when the cosine similarity between narrative embeddings exceeds 0.7 and the pair is 'manually reviewed.' The manual review process is not described: it is unclear who performed the review, whether it was the research team, how many pairs were added, and what the inter-reviewer agreement was. These judgments directly affect all R@K and NDCG numbers in Table 3, so the authors should report the number of added pairs, the review criteria, and agreement statistics, or at least provide a sensitivity analysis with and without the additional pairs.
minor comments (6)
  1. [Table 2 / Table 10] Tables 2 and 10 appear to report the same explanation results for the representative models; keeping both is redundant, and the paper should either merge them or clearly mark Table 2 as an excerpt of Table 10.
  2. [Table 2] The negative total score for Intern-S1-mini ( -2.234) should be explained, since the rubric sums positive dimensions and subtracts penalties; a brief note on how penalties can exceed positive credit would aid interpretation.
  3. [Section 3.3] The '85% agreement' between third-party annotators should specify whether this is simple percentage agreement or a chance-corrected measure such as Cohen's kappa, and the authors should clarify what exactly was agreed upon beyond inclusion decisions.
  4. [Abstract] The phrase 'implicit, non-linear, and rhetorically layered meanings' uses 'non-linear' where 'non-literal' seems intended; the paper should be checked for this potential terminology confusion.
  5. [Appendix D.1] The annotation section states that the main annotators were compensated 'as part of their research or project work'; it would be helpful to state the number of videos annotated by each annotator and whether any annotator skipped or flagged examples, since this bears on annotation consistency.
  6. [Figure 2] The ablation heatmaps in Figure 2 are dense and the color scale is not shown; adding a colorbar and highlighting the cells that deviated most from the ideal pattern would make the asymmetry claim easier to verify.

Circularity Check

0 steps flagged · score 2.0 of 10

Benchmark construction and evaluation are not circular; only minor self-citation and judge-family overlap appear, and they do not force the central claim.

full rationale

DrivelHub+ is a benchmark paper rather than a derivation chain. The headline result, that models 'can often describe what is shown, but still fail to infer what is meant,' is an empirical measurement produced by human-written explanations, an LLM-judge rubric, and bidirectional retrieval. No reported quantity is fitted from the data it is then used to predict. The retrieval relevance sets are anchored to the original video-narrative pairs, and the additional pairs are identified by EmbeddingGemma, which is not an evaluated retrieval model, then human-verified, so the evaluation is not forced by construction. The paper does rely on the authors' prior Drivelology concept (Wang et al., 2025b) and evaluates LCO-Omni (Xiao et al., 2026) co-authored by one of the present authors, but these are context and evaluated systems, not the evidence for the central gap. The judge Qwen3.6-35B-A3B belongs to the same model family as several evaluated systems, and the third-party agreement check in Section 3.3 covers inclusion decisions rather than the explanations themselves; these are validity and measurement risks, not circular reductions. The Limitations section itself acknowledges the task is inherently interpretive. The central claim would remain meaningful even if every self-citation were removed, so no significant circularity is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The evaluation rests on several assumed measurement instruments: human interpretations as ground truth, a single LLM judge, and an embedding threshold for relevance expansion. None is independently calibrated, which limits the strength of the numeric claims.

free parameters (2)
  • Cosine relevance threshold = 0.7
    Used in Section 4.3 to select additional relevant video-narrative pairs from EmbeddingGemma-300M nearest neighbours; changes the positive set and therefore retrieval metrics.
  • Judge rubric weights = Core 5, Rhetorical 3, Social 2, Grounding 2; penalties 3, 3, 2
    Hand-defined score ranges in Appendix E.2 shape all explanation total scores; no calibration against human ratings, so they are a free choice of the authors.
assumptions (4)
  • domain assumption Each selected video has a recoverable implicit pragmatic meaning independent of the annotator.
    Data collection criteria in Section 3.2 require an implicit meaning beyond literal description; if this is false for many videos, the benchmark measures annotator invention rather than model comprehension.
  • domain assumption The five research-team annotators' explanations constitute valid ground truth.
    Section 3.3 and Appendix D.1: the main annotators are research team members or close collaborators; only 100 videos were checked by third parties, with 85% agreement on inclusion decisions, not on explanation content.
  • ad hoc to paper Qwen3.6-35B-A3B provides valid judge scores without human validation.
    Section 4.2 and Appendix E.2 use this model as the sole judge; no human agreement study or bias analysis is reported, and the judge overlaps with the evaluated Qwen3.6 family.
  • domain assumption EmbeddingGemma-300M cosine similarity above 0.7 identifies narratives with the same implicit meaning.
    Section 4.3 uses this to expand relevance sets; the manual review mitigates but does not remove dependence on the embedding model's notion of similarity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reading Between the Frames: Interpreting Implicit and Non-literal Meaning in Social Media Videos." pith.science (2026). https://pith.science/paper/NS647DFB

@misc{pith2026260804939,
  author       = {Pith},
  title        = {Pith review of: Reading Between the Frames: Interpreting Implicit and Non-literal Meaning in Social Media Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NS647DFB}},
  note         = {Machine review of arXiv:2608.04939}
}
read the original abstract

Social media videos often communicate meanings that go beyond their visible actions, captions, or speech. A mundane clip may become humorous, ironic, or satire only through the interaction of multimodal cues and cultural context, making such content a difficult test case for video-language models. In this paper, we introduce \textit{DrivelHub+}, a benchmark for evaluating whether models can infer the implicit, non-linear, and rhetorically layered meanings of social media videos that appear nonsensical on the surface but convey deliberate pragmatic meanings. DrivelHub+ consists of 1,000 videos collected from social media, each annotated with a human-written implicit narrative explanation. Unlike conventional video understanding tasks focused on recognition or description, we present a benchmark that targets contextual multimodal reasoning. We evaluate current video-language models from two perspectives: explanation, where models must explain the pragmatic comprehension of a video in natural language; and representation, where we adapt reasoning-as-retrieval to test whether model representations align videos with their corresponding implicit narratives in both video-to-text and text-to-video retrieval. Our benchmark provides a diagnostic setting for measuring the gap between multimodal perception and pragmatic comprehension, asking whether current models can move beyond describing what is shown to inferring what is meant.

Figures

Figures reproduced from arXiv: 2608.04939 by the authors.

Figure 1
Figure 1. Example of a drivelological video where surface-level paper transformations encode historically associated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Input-stream ablation effects grouped by interpretive-evidence label. Rows are omni-modal models and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Human annotation guideline for implicit-meaning explanations. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Prompt used for explanation evaluation. that makes the video meaningful is important but secondary to recovering the implicit interpretation. Affective or Social Meaning and Grounding are assigned smaller ranges, [0,2], because they capture supporting aspects of the ex…
Figure 5
Figure 5. Figure 5: Prompt used for the vision-only generation setting, where the audio track is removed before inference. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Prompt used for the audio-only generation setting, where the visual stream is removed and the model [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Prompt used for Video LLM-as-a-judge evaluation. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: PCA of model-level judge-evaluation profiles. Each point represents a model using standardised rubric [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Side-by-side reasoning trace excerpts for the example in Figure [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Example of a drivelological video that requires on-screen wordplay but does not require cross-modal [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Example of a drivelological video requiring speech understanding, historical knowledge, and pragmatic [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: Example of a drivelological video requiring visual inference and frame-based humour. [PITH_FULL_IMAGE:figures/full_fig_p027_12.png]
Figure 13
Figure 13. Figure 13: Example of a drivelological video requiring the interaction of on-screen text and visual context. [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]
Figure 14
Figure 14. Figure 14: Example of a drivelological video requiring audio-visual pragmatic reasoning. [PITH_FULL_IMAGE:figures/full_fig_p028_14.png]
Figure 15
Figure 15. Figure 15: Example of a drivelological video requiring on-screen text, audio, and visual context. [PITH_FULL_IMAGE:figures/full_fig_p028_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 45 canonical work pages

  1. [1]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Drivel-ology: Challenging llms with interpreting nonsense with depth , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2025

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Egoschema: A diagnostic benchmark for very long-form video language understanding , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) , year=

    Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) , year=

    2024

  4. [4]

    arXiv preprint arXiv:2510.23727 , year=

    MUStReason: A Benchmark for Diagnosing Pragmatic Reasoning in Video-LMs for Multimodal Sarcasm Detection , author=. arXiv preprint arXiv:2510.23727 , year=

    work page arXiv

  5. [5]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Target-Augmented Shared Fusion-based Multimodal Sarcasm Explanation Generation , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  6. [6]

    arXiv preprint arXiv:2506.05425 , year=

    Siv-bench: A video benchmark for social interaction understanding and reasoning , author=. arXiv preprint arXiv:2506.05425 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2404.06347 , year=

    Rar-b: Reasoning as retrieval benchmark , author=. arXiv preprint arXiv:2404.06347 , year=

    arXiv

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Longvideobench: A benchmark for long-context interleaved video-language understanding , author=. Advances in Neural Information Processing Systems , volume=

Show all 77 references
  1. [9]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    WildVideo: Benchmarking LMMs for Understanding Video-Language Interaction , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  2. [10]

    arXiv preprint arXiv:2506.21742 , year=

    Implicitqa: Going beyond frames towards implicit video reasoning , author=. arXiv preprint arXiv:2506.21742 , year=

  3. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    MAVERIX: Multimodal Audio-Visual Evaluation and Recognition IndeX , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  4. [12]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    GODBench: A Benchmark for Multimodal Large Language Models in Video Comment Art , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  5. [13]

    2019 , publisher=

    Bad language , author=. 2019 , publisher=

    2019

  6. [14]

    2009 , publisher=

    On Bullshit , author=. 2009 , publisher=

    2009

  7. [15]

    Proceedings of the 18th International Natural Language Generation Conference , pages=

    Who’s laughing now? an overview of computational humour generation and explanation , author=. Proceedings of the 18th International Natural Language Generation Conference , pages=

  8. [16]

    ACM Transactions on Information Systems (TOIS) , volume=

    Cumulated gain-based evaluation of IR techniques , author=. ACM Transactions on Information Systems (TOIS) , volume=. 2002 , publisher=

    2002

  9. [17]

    Encyclopedia of database systems , pages=

    Mean reciprocal rank , author=. Encyclopedia of database systems , pages=. 2009 , publisher=

    2009

  10. [18]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  11. [19]

    Text summarization branches out , pages=

    Rouge: A package for automatic evaluation of summaries , author=. Text summarization branches out , pages=

  12. [20]

    VideoJudge: Bootstrapping Enables Scalable Supervision of

    Abdul Waheed and Zhen Wu and Dareen Safar Alharthi and Seungone Kim and Bhiksha Raj , booktitle=. VideoJudge: Bootstrapping Enables Scalable Supervision of. 2026 , url=

    2026

  13. [21]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Omnivid: A generative framework for universal video understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Nagrani, Arsha and Menon, Sachit and Iscen, Ahmet and Buch, Shyamal and Mehran, Ramin and Jha, Nilpa and Hauth, Anja and Zhu, Yukun and Vondrick, Carl and Sirotenko, Mikhail and Schmid, Cordelia and Weyand, Tobias , title =. Proceedings of the IEEE/CVF International Conference...

    2025

  15. [23]

    arXiv preprint arXiv:2503.06820 , year=

    Towards Fine-Grained Video Question Answering , author=. arXiv preprint arXiv:2503.06820 , year=

    arXiv

  16. [24]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    Tvqa+: Spatio-temporal grounding for video question answering , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  17. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Social-iq: A question answering benchmark for artificial social intelligence , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  18. [26]

    arXiv preprint arXiv:2406.04264 , volume=

    Mlvu: A comprehensive benchmark for multi-task long video understanding , author=. arXiv preprint arXiv:2406.04264 , volume=

    arXiv

  19. [27]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Lvbench: An extreme long video understanding benchmark , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  20. [28]

    arXiv preprint arXiv:2604.05015 , year=

    Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding , author=. arXiv preprint arXiv:2604.05015 , year=

    arXiv

  21. [29]

    CVPR , year=

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis , author=. CVPR , year=

  22. [30]

    NeurIPS Datasets and Benchmarks Track , year=

    Mme: A comprehensive evaluation benchmark for multimodal large language models , author=. NeurIPS Datasets and Benchmarks Track , year=

  23. [31]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , articleno =

    Farabi, Shafkat and Ranasinghe, Tharindu and Kanojia, Diptesh and Kong, Yu and Zampieri, Marcos , title =. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , articleno =. 2024 , isbn =. doi:10.24963/ijcai.2024/887 , abstract =

    2024 doi

  24. [32]

    Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages=

    Sarcasm Subtype-Specific Reasoning in Dialogue with Multimodal Cues Using Large Language Models , author=. Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages=

  25. [33]

    arXiv preprint arXiv:2604.16372 , year=

    CFMS: Towards Explainable and Fine-Grained Chinese Multimodal Sarcasm Detection Benchmark , author=. arXiv preprint arXiv:2604.16372 , year=

    arXiv

  26. [34]

    2026 , eprint=

    ViMU: Benchmarking Video Metaphorical Understanding , author=. 2026 , eprint=

    2026

  27. [35]

    arXiv preprint arXiv:2507.18552 , year=

    VideoMind: An Omni-Modal Video Dataset with Intent Grounding for Deep-Cognitive Video Understanding , author=. arXiv preprint arXiv:2507.18552 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2507.04909 , year=

    HumanVideo-MME: Benchmarking MLLMs for Human-Centric Video Understanding , author=. arXiv preprint arXiv:2507.04909 , year=

  29. [37]

    arXiv preprint arXiv:2405.08813 , year=

    CinePile: A Long Video Question Answering Dataset and Benchmark , author=. arXiv preprint arXiv:2405.08813 , year=

    arXiv

  30. [38]

    Workshop on Video-Language Models @ NeurIPS 2024 , year=

    CinePile: A Long Video Question Answering Dataset and Benchmark , author=. Workshop on Video-Language Models @ NeurIPS 2024 , year=

    2024

  31. [39]

    Advances in Neural Information Processing Systems , volume=

    Rextime: A benchmark suite for reasoning-across-time in videos , author=. Advances in Neural Information Processing Systems , volume=

  32. [40]

    2018 , publisher=

    Irony , author=. 2018 , publisher=

    2018

  33. [41]

    , author=

    A Pragmatic Study of Humor. , author=. Advances in Language and Literary Studies , volume=. 2016 , publisher=

    2016

  34. [42]

    2018 , publisher=

    Irony, deception and humour: Seeking the truth about overt and covert untruthfulness , author=. 2018 , publisher=

    2018

  35. [43]

    , author=

    A relevance-theoretic perspective on humorous irony and its failure. , author=. Humor: International Journal of Humor Research , volume=

  36. [44]

    Personality and individual differences , volume=

    Relations between humor styles and the Dark Triad traits of personality , author=. Personality and individual differences , volume=. 2010 , publisher=

    2010

  37. [45]

    arXiv preprint arXiv:2603.16859 , year=

    SocialOmni: Benchmarking Audio-Visual Social Interactivity in Omni Models , author=. arXiv preprint arXiv:2603.16859 , year=

    arXiv

  38. [46]

    arXiv e-prints , pages=

    Humanvbench: Exploring human-centric video understanding capabilities of mllms with synthetic benchmark data , author=. arXiv e-prints , pages=

  39. [47]

    2026 , eprint=

    GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning , author=. 2026 , eprint=

    2026

  40. [48]

    2025 , url=

    Holo2 - Open Foundation Models for Navigation and Computer Use Agents , author=. 2025 , url=

    2025

  41. [49]

    2025 , eprint=

    Intern-S1: A Scientific Multimodal Foundation Model , author=. 2025 , eprint=

    2025

  42. [50]

    2025 , eprint=

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency , author=. 2025 , eprint=

    2025

  43. [51]

    2024 , eprint=

    MiniCPM-V: A GPT-4V Level MLLM on Your Phone , author=. 2024 , eprint=

    2024

  44. [52]

    QVQ: To See the World with Wisdom , url =

  45. [53]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

    2025

  46. [54]

    2025 , eprint=

    Qwen3-VL Technical Report , author=. 2025 , eprint=

    2025

  47. [55]

    2025 , eprint=

    AVoCaDO: An Audiovisual Video Captioner Driven by Temporal Orchestration , author=. 2025 , eprint=

    2025

  48. [56]

    2025 , eprint=

    EchoInk-R1: Exploring Audio-Visual Reasoning in Multimodal LLMs via Reinforcement Learning , author=. 2025 , eprint=

    2025

  49. [57]

    2026 , eprint=

    WAVE: Learning Unified and Versatile Audio-Visual Embeddings with Multimodal LLM , author=. 2026 , eprint=

    2026

  50. [58]

    2026 , eprint=

    e5-omni: Explicit Cross-modal Alignment for Omni-modal Embeddings , author=. 2026 , eprint=

    2026

  51. [59]

    2025 , url=

    Octopus-Omni-Embed: Cost-Efficient Multi-Modal Embedding Model , author=. 2025 , url=

    2025

  52. [60]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Scaling Language-centric Omnimodal Representation Learning , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  53. [61]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  54. [62]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    LLMs as narcissistic evaluators: When ego inflates evaluation scores , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  55. [63]

    arXiv preprint arXiv:2605.08384 , year=

    jina-embeddings-v5-omni: Text-Geometry-Preserving Multimodal Embeddings via Frozen-Tower Composition , author=. arXiv preprint arXiv:2605.08384 , year=

    arXiv

  56. [64]

    2025 , eprint=

    EmbeddingGemma: Powerful and Lightweight Text Representations , author=. 2025 , eprint=

    2025

  57. [65]

    2025 , eprint=

    Thinking with Images for Multimodal Reasoning: Foundations, Methods, and Future Frontiers , author=. 2025 , eprint=

    2025

  58. [66]

    Towards Multimodal Sarcasm Detection (An Obviously Perfect Paper)

    Castro, Santiago and Hazarika, Devamanyu and P \'e rez-Rosas, Ver \'o nica and Zimmermann, Roger and Mihalcea, Rada and Poria, Soujanya. Towards Multimodal Sarcasm Detection (An Obviously Perfect Paper). Proceedings of the 57th Annual Meeting of the Association for Computation...

    2019 doi

  59. [67]

    UR-FUNNY: A multimodal language dataset for understanding humor , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  60. [68]

    SMILE : Multimodal Dataset for Understanding Laughter in Video with Language Models

    Hyun, Lee and Sung-Bin, Kim and Han, Seungju and Yu, Youngjae and Oh, Tae-Hyun. SMILE : Multimodal Dataset for Understanding Laughter in Video with Language Models. Findings of the Association for Computational Linguistics: NAACL 2024. 2024. doi:10.18653/v1/2024.findings-naacl.73

    2024 doi

  61. [69]

    2026 , eprint=

    v-HUB: A Benchmark for Video Humor Understanding from Vision and Sound , author=. 2026 , eprint=

    2026

  62. [70]

    arXiv preprint arXiv:2509.06771 , year=

    D-HUMOR: Dark Humor Understanding via Multimodal Open-ended Reasoning--A Benchmark Dataset and Method , author=. arXiv preprint arXiv:2509.06771 , year=

  63. [71]

    2025 , eprint=

    Qwen2.5-Omni Technical Report , author=. 2025 , eprint=

    2025

  64. [72]

    2025 , eprint=

    Qwen3-Omni Technical Report , author=. 2025 , eprint=

    2025

  65. [73]

    Target-Augmented Shared Fusion-based Multimodal Sarcasm Explanation Generation

    Goel, Palaash and Chauhan, Dushyant Singh and Akhtar, Md Shad. Target-Augmented Shared Fusion-based Multimodal Sarcasm Explanation Generation. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.472

    2025 doi

  66. [74]

    European Conference on Computer Vision (ECCV) , year =

    Binzhu Xie and Sicheng Zhang and Zitang Zhou and Bo Li and Yuanhan Zhang and Jack Hessel and Jingkang Yang and Ziwei Liu , title=. European Conference on Computer Vision (ECCV) , year =

  67. [75]

    ICCV , year =

    Long, Xinwei and Tian, Kai and Xu, Peng and Jia, Guoli and Li, Jingxuan and Yang, Sa and Shao, Yihua and Zhang, Kaiyan and Jiang, Che and Xu, Hao and Liu, Yang and Ma, Jiaheng and Zhou, Bowen , title =. ICCV , year =

  68. [76]

    Proceedings of the AAAI Conference on Artificial Intelligence , pages=

    VidEvent: A Large Dataset for Understanding Dynamic Evolution of Events in Videos , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=

  69. [77]

    Comparing Apples to Oranges: A Dataset & Analysis of LLM Humour Understanding from Traditional Puns to Topical Jokes

    Loakman, Tyler and Thorne, William and Lin, Chenghua. Comparing Apples to Oranges: A Dataset & Analysis of LLM Humour Understanding from Traditional Puns to Topical Jokes. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.finding...

    2025 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.