Pith. sign in

REVIEW 4 major objections 5 minor 41 references

IRONIC: Coherence-Aware Reasoning Chains for Multi-Modal Sarcasm Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A coherence-relation reasoning scaffold lets GPT-4o beat all prior zero-shot methods on multi-modal sarcasm detection.

desk verdict Promising prompting idea undercut by a missing ablation that separates coherence relations from richer task instructions; the state-of-the-art claim is unverifiable as presented. read the letter →

arxiv 2505.16258 v2 pith:DYBQV4G2 submitted 2025-05-22 cs.CL cs.AIcs.CV

Aashish Anantha Ramakrishnan , Aadarsh Anantha Ramakrishnan , Dongwon Lee This is my paper
classification cs.CLcs.AIcs.CV
keywords multi-modalsarcasmdetectioncoherencerelationsin-contextlearningchain-of-thoughtpromptingzero-shotmultimodallargelanguagemodelspragmaticsreasoningscaffolds
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

The paper argues that the reason chain-of-thought prompting underperforms on multi-modal sarcasm is that it does not mirror the pragmatic machinery humans use, so it proposes IRONIC, a two-step prompt scaffold built on Multi-modal Coherence Relations. In the first step the model names which of five coherence relations links image and text and explains why; in the second it uses that rationale, plus task-specific instructions, to decide whether sarcasm is present. The authors report that with GPT-4o this beats every previously published zero-shot result on the MMSD2.0 and RedEval datasets, and that it improves over zero-shot CoT in three of four model-dataset settings. The broader claim is that structured cognitive scaffolds help multimodal models reason about image-text incongruity where free-form step-by-step reasoning does not.

What carries the argument

The central object is the Multi-modal Coherence Relation, a discourse-level label describing how an image and its text are logically and pragmatically connected, grounded in the linguistic theory of discourse coherence. IRONIC uses five such relations—Insertion, Concretization, Projection, Restatement, and Extension—covering entity-level and scene-level referential, analogical, and pragmatic linkages. The machinery doing the work is the two-step prompt: first the model is asked to name the coherence relation and justify it, then a second prompt feeds that predicted relation and rationale back together with the original image-text pair and task-specific instructions to produce the final sarcasm label. The predicted relation functions as a cognitive scaffold that directs attention to the kind of incongruity at play, rather than letting the model free-associate its way through a generic chain of thought.

What would settle it

Run the IRONIC pipeline on the same datasets and models but replace the Coherence Relation prediction step with a placeholder step that asks only for a generic rationale, keeping the identical final prompt with its task-specific instructions; if GPT-4o's F1 on MMSD2.0 and RedEval stays near 0.7670 and 0.8422, the Coherence Relation step is not the cause of the improvement. A complementary check would feed randomly shuffled coherence-relation labels into the final prompt and observe whether performance collapses.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that explicit reasoning about the discourse-level relationship between an image and its caption—cast as one of five Multi-modal Coherence Relations (Insertion, Concretization, Projection, Restatement, Extension)—is a better reasoning chain for zero-shot sarcasm detection than generic step-by-step reasoning. In the proposed two-step framework IRONIC, a multimodal LLM first predicts the coherence relation and produces a supporting rationale, then makes the sarcasm decision conditioned on that rationale and on instructions that separate mockery and irony from simple humor or satire. The headline empirical finding is that GPT-4o with IRONIC reaches a weighted F1 of 0.7670 on MMSD2.0 and 0.8422 on RedEval, which the authors state beats all previous zero-shot MLLM-based approaches on both datasets; InternVL3-14B improves sharply on RedEval (0.8175 vs. 0.7428 for zero-shot CoT) but not on MMSD2.0, which the authors attribute to the smaller model's weaker capacity for generating high-quality intermediate rationales.

Load-bearing premise

The paper's central claim rests on the assumption that the performance gain of IRONIC over Zero-Shot CoT comes from the Coherence Relation prediction step and not from the extra task-specific instructions added to the final prompt, an assumption the paper does not test by ablation.

Editorial extensions

If this is right

  • If the result holds, task-specific fine-tuning is not required to reach state-of-the-art zero-shot sarcasm detection; a prompt-level cognitive scaffold suffices for a sufficiently capable model.
  • The five coherence relations provide a reusable vocabulary for image-text incongruity, so the same scaffold may transfer to other pragmatic tasks such as humor, irony, or hate-speech detection without retraining.
  • Larger multimodal models are predicted to benefit more from coherence-aware scaffolds because their internal reasoning fidelity is higher; the InternVL3 results on MMSD2.0 are the supporting observation.
  • The results imply that chain-of-thought prompting, as commonly applied, is not the right inductive bias for pragmatic multimodal understanding.
  • The distribution of predicted coherence relations (e.g., Projection and Restatement dominating MMSD2.0) suggests that sarcasm detection is often driven by analogical or explicit referential incongruity, which could guide dataset construction.

Reading between the lines

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

  • The paper leaves untested whether a direct ablation of the Coherence Relation step—removing CR prediction while keeping the identical final instructions—would preserve the gains; without that ablation the improved performance could be attributable to the extra task-specific wording alone.
  • The five CR labels are coarse; a natural extension would be to let models predict finer-grained relations or to automatically discover relation types from data, which could remove the need for hand-picked categories.
  • Because the same CR scaffold is task-independent, applying IRONIC to non-sarcastic figurative language (irony, satire, humor) on the same datasets would reveal whether the benefit is sarcasm-specific or general to pragmatic incongruity.
  • The reported per-sample CR distributions could serve as a diagnostic: if sarcastic samples cluster in particular relations, CR prediction could act as a cheap filter or explanation layer in deployed systems.
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

4 major / 5 minor

Summary. The paper proposes IRONIC, a two-step zero-shot prompting framework for multi-modal sarcasm detection. In the first step, an MLLM is asked to identify one of five Coherence Relations (Insertion, Concretization, Projection, Restatement, Extension) between the image and text and to produce a supporting rationale. In the second step, the model uses this analysis, together with an enriched task instruction, to classify the pair as sarcastic or non-sarcastic. Experiments with GPT-4o and InternVL3-14B on MMSD2.0 and RedEval compare IRONIC against Zero-Shot, Zero-Shot CoT, and S3 CoT baselines. The paper reports that GPT-4o with IRONIC achieves state-of-the-art zero-shot results, while InternVL3 benefits less consistently across datasets.

Significance. If the causal claim holds, the paper would make a useful contribution by showing that discourse-coherence structure improves MLLM reasoning for a pragmatic task such as sarcasm detection. The release of code and prompt templates supports reproducibility, and the evaluation across two datasets and two model families is a reasonable starting point. However, the significance is currently undermined by an untested confound between the coherence-relation step and the added task-specific instruction, by the absence of direct numerical comparison with the cited state-of-the-art systems, and by the lack of statistical reliability measures. These issues must be resolved before the contribution can be properly assessed.

major comments (4)
  1. [Section 4 and Appendix D] The reported improvements of IRONIC over Zero-Shot CoT are confounded. The IRONIC user prompt differs from the Zero-Shot CoT prompt in two material ways: it asks the model to identify the most appropriate coherence relation, and it adds a task-specific instruction defining multi-modal sarcasm and directing the model to distinguish humor/satire from mockery/irony. The Zero-Shot CoT final prompt only asks for a 0/1 judgment from the generated rationale. Because the paper never ablates the coherence-relation request while retaining the added instructions (or vice versa), the observed F1 gains with GPT-4o (3.23% on MMSD2.0 and 9.70% on RedEval) cannot be attributed to the coherence-relation step; they could be driven entirely by the improved task description. An ablation with conditions that vary these two components independently is necessary to support the paper's central causal claim.
  2. [Section 4, 'IRONIC Improves over CoT Baselines'] The statement that IRONIC 'beats all previous state-of-the-art Zero-Shot approaches' cites Zhang et al. (2025b) (Commander-GPT) and Tang et al. (2024), but Table 1 contains no results for either system. The claim is therefore not supported by the evidence presented in this manuscript. The authors should either include a direct comparison with these systems under the same evaluation protocol, with numbers in the table, or temper the claim to what the presented experiments can actually support.
  3. [Table 1 and Appendix C] All experimental results appear to be based on a single run per condition. Setting temperature to 0 or do_sample=False reduces but does not eliminate variance for API-based models, and the observed margins are small (e.g., 0.7659 vs 0.7468 for GPT-4o on MMSD2.0). The paper should report multiple runs with standard deviations and, where feasible, significance tests, so that the reader can judge whether the differences are reliable.
  4. [Table 1, InternVL3 rows] The InternVL3 results undermine the generality of the central claim. On MMSD2.0, IRONIC achieves an F1 of 0.7032, below Zero-Shot CoT's 0.7400 and Zero-Shot's 0.7322; on RedEval, IRONIC's 0.8175 is below the Zero-Shot baseline of 0.8309. This pattern is consistent with the hypothesis that the added task instructions, rather than the coherence-relation step, drive the observed gains on GPT-4o. The proposed ablation should explicitly test this interaction, since the current text does not resolve it.
minor comments (5)
  1. [Section 3, Datasets] The Datasets subsection cites 'RedEval (Tong et al., 2024)', while the Introduction and the reference list attribute the dataset to Tang et al. (2024). Please correct the citation and ensure the reference entry matches.
  2. [Abstract and Section 2] The abstract and introduction describe the Coherence Relations as covering referential, analogical, and pragmatic image-text linkages, but the Methodology section (Section 2) at one point mentions only referential and pragmatic reasoning types. Please make the terminology consistent.
  3. [Appendix D, IRONIC Prompt] The system message says 'analyze and predict the most applicable Coherence Relations' while the user prompt asks the model to 'identify the most appropriate coherence relation'; please align the wording across the two parts of the prompt.
  4. [Table 1] The table header uses 'GPT4o' while the rest of the paper uses 'GPT-4o'; please standardize the model name.
  5. [Section 5, Conclusion] The sentence 'Our experiments also investigate the distinction between implicit pattern matching and explicit reasoning' overstates the evidence: the experiments compare prompting strategies but do not directly measure or manipulate implicit versus explicit reasoning processes.

Circularity Check

0 steps flagged · score 2.0 of 10

IRONIC's reported gains rest on an empirical zero-shot comparison against external held-out test sets, and the self-cited coherence-relation framework is corroborated by external sources (Hobbs 1978; Alikhani et al. 2020; Xu et al. 2022); the missing CR-only ablation is a causal-attribution gap, not a circular derivation.

full rationale

The core result is an empirical measurement, not a derivation: GPT-4o with IRONIC reaches 0.7670 and 0.8422 weighted F1 versus Zero-Shot CoT's 0.7430 and 0.7677 on the externally constructed MMSD2.0 and RedEval test sets, with no parameter fitting and no equation making outputs equivalent to inputs by construction, so no self-definitional or fitted-input pattern is present. The coherence-relation taxonomy is traced to Hobbs (1978), Alikhani et al. (2020), and Xu et al. (2022), all external to the authors; the self-citations (Anantha Ramakrishnan et al., 2025a/b, CORDIAL and RONA) are invoked alongside those external references for the same premises, and the other motivational claims are dual-cited with Zhang et al. (2025a) and Jia et al. (2025), so per the review rules they are corroborated evidence, not load-bearing self-citation. The flagged weakness, located in the prompt templates of Appendix D versus the Zero-Shot CoT template, is that the IRONIC final prompt differs in two respects at once: it asks the model to 'identify the most appropriate coherence relation' and adds task-specific instructions ('Content where both image and text together create a sarcastic effect is considered multi-modal sarcasm... Clearly distinguish simple humor and satire which is not sarcastic from mockery and irony which is sarcastic'); since no ablation removes the coherence-relation request while retaining the added instructions, the paper's implicit causal claim that coherence reasoning drives the improvement is under-supported. This is an untested attribution, however, not a circular reduction, because the model's 0/1 outputs are judged against held-out human labels and the paper never reads the sarcasm decision off the predicted coherence relation. The InternVL3-14B results reinforce the non-circular reading: IRONIC is worse than Zero-Shot CoT on MMSD2.0 (0.7032 vs 0.7400 F1), and the paper openly states that 'the introduction of reasoning chains, both general and coherence-aware, does not provide meaningful performance improvements' for that model, scoping its state-of-the-art claim to GPT-4o. A result forced by construction would not exhibit such model dependence. Overall, the paper is self-contained against external benchmarks, and the only circularity-adjacent signals are minor, non-load-bearing self-citations.

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

The framework introduces no fitted parameters; its load-bearing assumptions are the adequacy of the five chosen Coherence Relations and the usefulness of model-generated CR rationales for the final classification. The absence of an ablation makes these assumptions untested.

assumptions (4)
  • ad hoc to paper The five Coherence Relations (Insertion, Concretization, Projection, Restatement, Extension) adequately cover referential, analogical, and pragmatic image-text relations relevant to sarcasm.
    Chosen by the authors from prior work and inserted into the system prompt; no validation or comparison with alternative relation taxonomies is provided (Section 2 and Appendix D).
  • domain assumption MLLM-generated Coherence Relation predictions and rationales are accurate enough to improve subsequent sarcasm classification.
    The framework relies on the quality of the Relation Extraction step, but the paper does not evaluate CR prediction accuracy or rationale quality, and the final prompt is designed to tolerate upstream errors (Section 2 and Limitations).
  • domain assumption The MMSD2.0 and RedEval test sets are representative benchmarks for zero-shot multi-modal sarcasm detection and are directly comparable across prompting methods.
    Used without further validation; no discussion of label noise or domain shift beyond dataset construction notes (Section 3).
  • domain assumption A single decoding at temperature 0 with seed 42 is representative of GPT-4o and InternVL3 performance on these tasks.
    No repeated runs or variance reporting (Appendix C).

how reviews work

0 comments
Cite this review

Pith. "Pith review of IRONIC: Coherence-Aware Reasoning Chains for Multi-Modal Sarcasm Detection." pith.science (2026). https://pith.science/paper/DYBQV4G2

@misc{pith2026250516258,
  author       = {Pith},
  title        = {Pith review of: IRONIC: Coherence-Aware Reasoning Chains for Multi-Modal Sarcasm Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYBQV4G2}},
  note         = {Machine review of arXiv:2505.16258}
}
read the original abstract

Interpreting figurative language such as sarcasm across multi-modal inputs presents unique challenges, often requiring task-specific fine-tuning and extensive reasoning steps. However, current Chain-of-Thought approaches do not efficiently leverage the same cognitive processes that enable humans to identify sarcasm. We present IRONIC, an in-context learning framework that leverages Multi-modal Coherence Relations to analyze referential, analogical and pragmatic image-text linkages. Our experiments show that IRONIC achieves state-of-the-art performance on zero-shot Multi-modal Sarcasm Detection across different baselines. This demonstrates the need for incorporating linguistic and cognitive insights into the design of multi-modal reasoning strategies. Our code is available at: https://github.com/aashish2000/IRONIC

Figures

Figures reproduced from arXiv: 2505.16258 by the authors.

Figure 1
Figure 1. Overview of the proposed IRONIC framework strategies (Wang & Zhao, 2024; Lee et al., 2025), we propose IRONIC (explIcit Reasoning with cOherence relatioNs as cognItive struCtures). IRONIC leverages multi-modal Co￾herence Relations (Anantha Ramakrishnan et al., 2025a; Alikhani et al., 2020) to provide a structured reasoning pathway for MLLMs to follow while identifying image-text incon￾gruencies. Coherence Relations … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 35 canonical work pages

  1. [1]

    Cross-modal coherence modeling for caption generation

    Malihe Alikhani, Piyush Sharma, Shengjie Li, Radu Soricut, and Matthew Stone. Cross-modal coherence modeling for caption generation. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.\ 6525--6535, Stroudsburg, PA, USA, 2020. Association for ...

    work page 2020

  2. [2]

    Cross-modal coherence for text-to-image retrieval

    Malihe Alikhani, Fangda Han, Hareesh Ravi, Mubbasir Kapadia, Vladimir Pavlovic, and Matthew Stone. Cross-modal coherence for text-to-image retrieval. Proc. Conf. AAAI Artif. Intell., 36 0 (10): 0 10427--10435, June 2022

    work page 2022

  3. [3]

    ANNA : Abstractive text-to-image synthesis with filtered news captions

    Aashish Anantha Ramakrishnan, Sharon X Huang, and Dongwon Lee. ANNA : Abstractive text-to-image synthesis with filtered news captions. In The Third Workshop on Advances in Language and Vision Research. Association for Computational Linguistics, 2024

    work page 2024

  4. [4]

    CORDIAL : Can multimodal large language models effectively understand coherence relationships? arXiv [cs.CL], February 2025 a

    Aashish Anantha Ramakrishnan, Aadarsh Anantha Ramakrishnan, and Lee Dongwon. CORDIAL : Can multimodal large language models effectively understand coherence relationships? arXiv [cs.CL], February 2025 a

    work page 2025

  5. [5]

    RONA : Pragmatically diverse image captioning with coherence relations

    Aashish Anantha Ramakrishnan, Aadarsh Anantha Ramakrishnan, and Dongwon Lee. RONA : Pragmatically diverse image captioning with coherence relations. In Proceedings of the Fourth Workshop on Intelligent and Interactive Writing Assistants (In2Writing 2025), pp.\ 74--86. Association for Computational Linguistics, May 2025 b

    work page 2025

  6. [6]

    Multi-modal sarcasm detection in twitter with hierarchical fusion model

    Yitao Cai, Huiyu Cai, and Xiaojun Wan. Multi-modal sarcasm detection in twitter with hierarchical fusion model. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.\ 2506--2515, Stroudsburg, PA, USA, 2019. Association for Computational Linguistics

    work page 2019

  7. [7]

    Sarcasm, pretense, and the semantics/pragmatics distinction

    Elisabeth Camp. Sarcasm, pretense, and the semantics/pragmatics distinction. Nous, 46 0 (4): 0 587--634, December 2012

    work page 2012

  8. [8]

    Towards multimodal sarcasm detection (an \_obviously\_ perfect paper)

    Santiago Castro, Devamanyu Hazarika, Verónica Pérez-Rosas, Roger Zimmermann, Rada Mihalcea, and Soujanya Poria. Towards multimodal sarcasm detection (an \_obviously\_ perfect paper). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.\ 4619--4629, Stroudsburg, PA, USA, 2019. Association for Computational Linguistics

    work page 2019

Show all 41 references
  1. [9]

    Sarcasm detection on facebook: a supervised learning approach

    Dipto Das and Anthony J Clark. Sarcasm detection on facebook: a supervised learning approach. In Proceedings of the 20th International Conference on Multimodal Interaction: Adjunct, New York, NY, USA, October 2018. ACM

    2018

  2. [10]

    Why is discourse coherent?, volume 176

    Jerry R Hobbs. Why is discourse coherent?, volume 176. SRI International Menlo Park, CA, 1978

    1978

  3. [11]

    Exploring and evaluating multimodal knowledge reasoning consistency of multimodal large language models

    Boyu Jia, Junzhe Zhang, Huixuan Zhang, and Xiaojun Wan. Exploring and evaluating multimodal knowledge reasoning consistency of multimodal large language models. arXiv [cs.CL], March 2025

    2025

  4. [12]

    Debiasing multimodal sarcasm detection with contrastive learning

    Mengzhao Jia, Can Xie, and Liqiang Jing. Debiasing multimodal sarcasm detection with contrastive learning. Proc. Conf. AAAI Artif. Intell., 38 0 (16): 0 18354--18362, March 2024

    2024

  5. [13]

    When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues

    Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tanmoy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: ...

    2022

  6. [14]

    Pragmatic metacognitive prompting improves LLM performance on sarcasm detection

    Joshua Lee, Wyatt Fong, Alexander Le, Sur Shah, Kevin Han, and Kevin Zhu. Pragmatic metacognitive prompting improves LLM performance on sarcasm detection. In Proceedings of the 1st Workshop on Computational Humor (CHum), pp.\ 63--70, 2025

    2025

  7. [15]

    Multi-modal sarcasm detection via cross-modal graph convolutional network

    Bin Liang, Chenwei Lou, Xiang Li, Min Yang, Lin Gui, Yulan He, Wenjie Pei, and Ruifeng Xu. Multi-modal sarcasm detection via cross-modal graph convolutional network. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pape...

    2022

  8. [16]

    FuseDream : Training-free text-to-image generation with improved CLIP+GAN space optimization

    Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Hao Su, and Qiang Liu. FuseDream : Training-free text-to-image generation with improved CLIP+GAN space optimization. arXiv:2112.01573 [cs], December 2021

    2021 arXiv

  9. [17]

    Diverse image captioning with context-object split latent spaces

    Shweta Mahajan and Stefan Roth. Diverse image captioning with context-object split latent spaces. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates Inc

    2020

  10. [18]

    GPT - 4o system card

    OpenAI , Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, A J Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino...

    2024

  11. [19]

    Modeling intra and inter-modality incongruity for multi-modal sarcasm detection

    Hongliang Pan, Zheng Lin, Peng Fu, Yatao Qi, and Weiping Wang. Modeling intra and inter-modality incongruity for multi-modal sarcasm detection. In Findings of the Association for Computational Linguistics: EMNLP 2020, pp.\ 1383--1392, Stroudsburg, PA, USA, November 2020. Assoc...

    2020

  12. [20]

    Opinion mining and sentiment analysis

    Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Found. Trends® Inf. Retr., 2 0 (1-2): 0 1--135, July 2008

    2008

  13. [21]

    MMSD2 .0: Towards a reliable multi-modal sarcasm detection system

    Libo Qin, Shijue Huang, Qiguang Chen, Chenran Cai, Yudi Zhang, Bin Liang, Wanxiang Che, and Ruifeng Xu. MMSD2 .0: Towards a reliable multi-modal sarcasm detection system. In Findings of the Association for Computational Linguistics: ACL 2023, pp.\ 10834--10845, Stroudsburg, PA...

    2023

  14. [22]

    Learning transferable visual models from natural language supervision

    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. Learning transferable visual models from natural language supervision. arXiv:2103.00020 [cs...

    2021 arXiv

  15. [23]

    A multimodal corpus for emotion recognition in sarcasm

    Anupama Ray, Shubham Mishra, Apoorva Nunna, and Pushpak Bhattacharyya. A multimodal corpus for emotion recognition in sarcasm. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pp.\ 6992--7003, 2022

    2022

  16. [24]

    Detecting sarcasm in multimodal social platforms

    Rossano Schifanella, Paloma de Juan, Joel Tetreault, and Liangliang Cao. Detecting sarcasm in multimodal social platforms. In Proceedings of the 24th ACM international conference on Multimedia, New York, NY, USA, October 2016. ACM

    2016

  17. [25]

    Using the image-text relationship to improve multimodal disaster tweet classification

    Tiberiu Sosea, Iustin Sirbu, Cornelia Caragea, Doina Caragea, and Traian Rebedea. Using the image-text relationship to improve multimodal disaster tweet classification. Int Conf Inf Syst Crisis Response Manag, pp.\ 691--704, 2021

    2021

  18. [26]

    Leveraging generative large language models with visual instruction and demonstration retrieval for multimodal sarcasm detection

    Binghao Tang, Boda Lin, Haolong Yan, and Si Li. Leveraging generative large language models with visual instruction and demonstration retrieval for multimodal sarcasm detection. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computat...

    2024

  19. [27]

    Eyes wide shut? exploring the visual shortcomings of MultiModal LLMs

    Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of MultiModal LLMs . Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., pp.\ 9568--9578, January 2024

    2024

  20. [28]

    Categorizing and inferring the relationship between the text and image of twitter posts

    Alakananda Vempala and Daniel Preoţiuc-Pietro. Categorizing and inferring the relationship between the text and image of twitter posts. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.\ 2830--2840, Stroudsburg, PA, USA, July 2019....

    2019

  21. [29]

    S ^ 3 agent: Unlocking the power of VLLM for zero-shot multi-modal sarcasm detection

    Peng Wang, Yongheng Zhang, Hao Fei, Qiguang Chen, Yukai Wang, Jiasheng Si, Wenpeng Lu, Min Li, and Libo Qin. S ^ 3 agent: Unlocking the power of VLLM for zero-shot multi-modal sarcasm detection. ACM Trans. Multimed. Comput. Commun. Appl., August 2024

    2024

  22. [30]

    Metacognitive prompting improves understanding in large language models

    Yuqing Wang and Yun Zhao. Metacognitive prompting improves understanding in large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp.\ 19...

    2024

  23. [31]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and Others . Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  24. [32]

    Modeling incongruity between modalities for multimodal sarcasm detection

    Yang Wu, Yanyan Zhao, Xin Lu, Bing Qin, Yin Wu, Jian Sheng, and Jinlong Li. Modeling incongruity between modalities for multimodal sarcasm detection. IEEE Multimed., 28 0 (2): 0 86--95, April 2021

    2021

  25. [33]

    Understanding social media cross-modality discourse in linguistic space

    Chunpu Xu, Hanzhuo Tan, Jing Li, and Piji Li. Understanding social media cross-modality discourse in linguistic space. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (eds.), Findings of the Association for Computational Linguistics: EMNLP 2022, pp.\ 2459--2471, Stroudsburg...

    2022

  26. [34]

    Is sarcasm detection a step-by-step reasoning process in large language models? Proc

    Ben Yao, Yazhou Zhang, Qiuchi Li, and Jing Qin. Is sarcasm detection a step-by-step reasoning process in large language models? Proc. Conf. AAAI Artif. Intell., 39 0 (24): 0 25651--25659, April 2025

    2025

  27. [35]

    RATT : A thought structure for coherent and correct LLM reasoning

    Jinghan Zhang, Xiting Wang, Weijieying Ren, Lu Jiang, Dongjie Wang, and Kunpeng Liu. RATT : A thought structure for coherent and correct LLM reasoning. Proc. Conf. AAAI Artif. Intell., 39 0 (25): 0 26733--26741, April 2025 a

    2025

  28. [36]

    Commander- GPT : Fully unleashing the sarcasm detection capability of multi-modal large language models

    Yazhou Zhang, Chunwang Zou, Bo Wang, and Jing Qin. Commander- GPT : Fully unleashing the sarcasm detection capability of multi-modal large language models. arXiv [cs.CL], March 2025 b

    2025

  29. [37]

    InternVL3 : Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, Hao Li, Jiahao Wang, Dengnian Chen, Songze Li, Yinan He, Ta...

    2025

  30. [38]

    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 gl...

  31. [39]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  32. [40]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  33. [41]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.