Pith. sign in

REVIEW 3 major objections 6 minor 64 references

Symmetric charged prompts cancel LLM over-prediction of sarcasm, and late acoustic fusion rescues weak models without fine-tuning any backbone.

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 · grok-4.5

2026-07-14 07:02 UTC pith:JDNTHOWH

load-bearing objection Solid systems paper: symmetric charge ensembling plus late acoustic rescue actually moves zero-shot sarcasm numbers, with a clean English–Chinese prosody split; the additive-bias math is a story, not a proof. the 3 major comments →

arxiv 2607.11102 v1 pith:JDNTHOWH submitted 2026-07-13 cs.SD

CHARM: Charge Calibration and Acoustic Rescue for LLM-based Multimodal Sarcasm Detection

Qiyang Sun , Yi Chang , Yupei Li , Xi Shao , Zixing Zhang , Bj\"orn W. Schuller This is my paper
classification cs.SD
keywords Multimodal Sarcasm DetectionLarge Language ModelsZero-Shot LearningCharge CalibrationAcoustic Late FusionCross-Cultural AnalysisProsodic DecouplingExplainability
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.

Instruction-tuned language models systematically over-call sarcasm in zero-shot settings because alignment rewards agreeing with the question's framing. CHARM treats that bias as a signed charge and cancels it by querying each utterance with five symmetrically opposing prompts, then majority-voting the answers into a single calibrated text score. When text alone still collapses—especially on weaker models or imbalanced Chinese dialogue—a second stage fuses that score with raw pitch and loudness statistics plus multi-class auditory-perception probes from a frozen audio-language model, through a shallow classifier that can down-weight saturated text votes. The result is the highest reported training-free text-only Macro-F1 on the English MUStARD benchmark and large rescues of weak backbones on Chinese CMMA, plus evidence that low-level prosody fails to transfer across languages while high-level perceptual abstractions remain useful. A sympathetic reader cares because the method keeps billion-parameter models frozen, works behind commercial APIs, and yields an explainable, cross-lingual detector that does not require scarce sarcasm-annotated audio.

Core claim

Zero-shot instruction-tuned LLMs over-predict the sarcastic class across the capability spectrum; Bidirectional Charge Calibration (BiCAL) cancels that directional bias by construction with a symmetric prompt axis, recovering an unbiased pragmatic vote, while Acoustic Late-Fusion Rescue (ALFR) then restores performance on weak or culturally mismatched backbones by fusing the calibrated vote with openSMILE prosody and Omni perception probes through a shallow head—without fine-tuning any foundation model.

What carries the argument

Bidirectional Charge Calibration (BiCAL): five charge-typed prompts spanning +2 to −2 induce opposing directional biases that sum to zero; nested majority voting over templates and seeds yields a 0–15 confidence score. Acoustic Late-Fusion Rescue (ALFR): a 27-dimensional vector (1 text vote + 6 speaker-normalised openSMILE features + 20 one-hot Omni probes) is classified by a lightweight head that re-weights toward acoustics when text saturates.

Load-bearing premise

The paper treats prompt-induced sycophancy as an additive, roughly symmetric bias on the model’s log-odds so that five hand-designed opposing prompts cancel each other and leave only the true sarcasm signal.

What would settle it

Replace the five charged templates with five neutral paraphrases of identical length and sampling budget; if Macro-F1 no longer rises above the single-prompt baseline on the same backbone and data, the gain was ensemble smoothing rather than charge cancellation.

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

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes CHARM, a two-stage framework for multimodal sarcasm detection. Bidirectional Charge Calibration (BiCAL) ensembles five symmetrically charged zero-shot prompts so that prompt-induced positivity bias is intended to cancel, recovering a calibrated text vote without fine-tuning. Acoustic Late-Fusion Rescue (ALFR) then concatenates that vote with speaker-normalised openSMILE prosody and multi-class Omni perception probes into a 27-D vector scored by a shallow classifier. Across seven LLM backbones on English MUStARD and Chinese CMMA, BiCAL reports consistent Macro-F1 gains and a peak zero-shot text-only Macro-F1 of 0.787; ALFR further rescues weak backbones by up to +0.382 Macro-F1. Stouffer meta-analyses, a random-paraphrase control, feature ablations, ensemble-size sweeps, and impurity-based routing analyses support the empirical claims and a cross-cultural finding that low-level prosody transfers poorly while high-level perception probes remain robust.

Significance. If the empirical results hold under the stated protocols, CHARM is a useful, practical contribution to zero-shot and lightly supervised multimodal sarcasm detection. Strengths include evaluation on seven backbones spanning open and API models, bilingual benchmarks with severe class imbalance on CMMA, a polarity-vs-paraphrase control (Table IV), feature-block ablations (Table V), cost–accuracy ensemble sweeps (Table VI), bootstrap/Stouffer significance testing, and impurity-based explainability of the late-fusion head. The cross-cultural prosodic decoupling analysis is of independent interest and is tied to prior psycholinguistic work. Public code and fixed library hyperparameters for the fusion heads further support reproducibility. The work sits cleanly between pure prompting and full multimodal fine-tuning, which is valuable given scarce sarcasm audio labels and closed API models.

major comments (3)
  1. [Section III-C, Eqs. (3)–(5)] Section III-C, Eqs. (3)–(5): The central mechanistic claim that directional biases “cancel by construction” rests on decomposing latent log-odds as L(u,c,p)=f(u,c)+δ(p) with a hand-designed five-template axis enforcing Σδ(p)=0. This requires δ to be approximately additive, content-independent, and equal-magnitude opposite for ±k charges. The manuscript does not report per-charge positive rates, content-conditioned bias estimates, or a direct zero-sum check on the stronger backbones that set the headline 0.787 result. Without such diagnostics, the abstract/intro language overstates a design intention as a proven cancellation property; gains may partly be ordinary multi-prompt self-consistency. Please either (i) add per-charge and content-dependence diagnostics across backbones, or (ii) substantially soften “cancel by construction” / “unbiased pragmatic signal” claims to match what is actu
  2. [Table IV; Table II (DeepSeek-V4-Pro)] Table IV and the 0.787 headline: The random-paraphrase control that isolates polarity from mere prompt/sampling diversity is reported only for Qwen2.5-32B-Instruct on MUStARD (BiCAL 0.713 vs RandPara 0.692). The strongest text-only claim is DeepSeek-V4-Pro at Macro-F1 0.787 (Table II). A single-backbone control is insufficient to underwrite the mechanism story for the peak result. Extend the paraphrase (or charge-scrambled) control to at least the top-performing backbones, or qualify the mechanism claim so that it is not carried by the untested models.
  3. [Abstract; Section I contributions; Eq. (2)] Abstract vs. Stage-2 design: The abstract calls CHARM “a training-free framework,” yet ALFR trains a shallow classifier under stratified CV with class re-weighting (Section III-D, IV-C, Appendix A). The body correctly distinguishes zero-shot BiCAL from lightly supervised ALFR, but the abstract and some framing sentences blur this. Please make the training regime explicit in the abstract and contribution bullets (e.g., training-free BiCAL; frozen-backbone light-sup ALFR) so the supervision claim matches the method.
minor comments (6)
  1. [Table III; Section V-C] Table III comparisons are acknowledged as heterogeneous (supervision, modality, splits). Still, the “highest reported zero-shot text-only Macro-F1” claim would be clearer if the table explicitly marked evaluation subset/split differences for CMMA (audio-available English-prompt subset) and MUStARD full-set vs prior splits.
  2. [Figure 1; Eq. (2)] Figure 1 and Eq. (2) use slightly inconsistent stage naming (Mode A/B vs Stage 1/2). Align terminology throughout.
  3. [Eq. (2); Section III-C] The fixed decision rule v>7 on a 0–15 vote sum is a free threshold; a short sensitivity note (or calibration of the threshold on a tiny held-out slice for Stage 1) would help readers assess robustness.
  4. [Eq. (8); Section III-D] Speaker-normalisation cascade (Eq. 8) uses n_min=2 (MUStARD) vs 5 (CMMA). Briefly justify these cutoffs and report how often each pool level is used.
  5. [Section III; Appendix E] Minor typography: “BICAL” vs “BiCAL” appears inconsistently in a few places (e.g., early methodology prose); standardise on BiCAL. Also check arXiv id / model-card naming consistency for frontier APIs.
  6. [Section VI; Figure 3 (P3)] Limitations correctly note single Omni probe model and English–Chinese only coverage; a one-sentence note on whether P3 (tone–text match) could leak transcript content into the “acoustic” block would strengthen the modality-separation claim.

Circularity Check

1 steps flagged

No load-bearing circular derivation: BiCAL's 'cancel by construction' is intentional prompt design under an additive-bias ansatz; Macro-F1 and ALFR gains are measured against external labels and CV folds.

specific steps
  1. self definitional [Abstract; §III-C Eqs. (3)–(5)]
    "the induced directional biases cancel by construction, and a simple aggregation recovers an unbiased pragmatic signal. ... Due to human preference alignment constraints, the predictive distribution suffers from directional bias. We decompose this latent log odds function into two main components L(u, c, p) = f(u, c) + δ(p), ... our symmetric prompt layout enforces a zero-sum equilibrium over the bias space ∑_{p∈C} δ(p) = 0."

    Zero-sum cancellation is imposed by authoring five hand-symmetric charge templates and assuming additive, content-independent δ(p); under that model the 'unbiased' vote is the ensemble procedure by definition. This is design-level circularity of the mechanism claim only. It does not force the external Macro-F1 against MUStARD/CMMA labels, which remain independent measurements.

full rationale

CHARM is an empirical methods paper. BiCAL defines a five-template symmetric charge axis and, under the modeling assumption L(u,c,p)=f(u,c)+δ(p) with hand-enforced Σδ(p)=0 (Eqs. 3–5), aggregates votes; the phrase 'cancel by construction' describes that design choice, not a fitted quantity renamed as a prediction. The headline Macro-F1 numbers (0.787 on MUStARD, ALFR lifts on CMMA) are evaluated against ground-truth sarcasm labels and DZS baselines, with bootstrap/Stouffer tests and a random-paraphrase control (Table IV) that is not forced by the charge definition. ALFR trains only a shallow head under stratified CV and reports held-out metrics; feature-importance and cross-cultural Cohen-d analyses are post-hoc diagnostics, not self-fulfilling targets. There is no uniqueness theorem imported from overlapping authors, no fitted parameter re-sold as an independent prediction of a closely related quantity, and no renaming of a known closed-form result. The additive, content-independent δ(p) model is an untested modeling assumption (correctness risk), not circularity of the reported scores. Score 1 reflects only the mild design-level 'by construction' wording, not a reduction of the central empirical claims.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

Central claims rest on a domain model of additive prompt bias, hand-crafted charge templates, fixed decoding and fusion hyperparameters, and the assumption that frozen Omni multi-class probes plus openSMILE functionals carry sarcasm-relevant signal. No new physical entities; free knobs are methodological. Empirical support is external benchmarks, not circular definitions.

free parameters (5)
  • BiCAL charge templates (5 hand-authored System/User tails on axis −2…+2)
    Wording of expert/skeptic roles is designer-chosen; polarity—not a fitted continuous parameter—drives the method, but results depend on this specific continuum.
  • Decoding temperature T=0.5, top-p=0.95, 3 samples × 3 seeds
    Fixed sampling profile that defines the 15-vote ensemble; different T would change vote variance and Macro-F1.
  • Speaker-normalization n_min (2 MUStARD / 5 CMMA) and z-clip [−5,+5]
    Dataset-specific thresholds for hierarchical openSMILE z-scoring; chosen by authors, not learned from sarcasm labels but still free design knobs.
  • Late-fusion classifier family and fixed library hyperparameters
    Per-backbone/dataset choice among 10 heads (AdaBoost on MUStARD, Soft Voting on CMMA; C, n_est, η, class_weight, scale_pos_weight=7.8, etc.) is selected on CV performance.
  • Decision threshold v>7 on 0–15 vote sum
    Hard majority cut for Stage-1 text-only decisions; midpoint of the discrete scale.
axioms (4)
  • domain assumption Instruction-tuned LLMs exhibit directional sycophancy/positivity bias on sarcasm queries that can be decomposed as L(u,c,p)=f(u,c)+δ(p) with opposing charges approximately summing to zero.
    Section III-C Eqs. 3–5; motivated by RLHF literature but not measured at logit level in this paper.
  • standard math Logistic mapping is strictly monotonic and centrosymmetric, so symmetric bias in log-odds cancels under majority aggregation.
    Used to justify bias cancellation without logit access.
  • domain assumption Low-level eGeMAPS functionals and frozen Qwen2.5-Omni multi-class probes carry usable sarcasm-related acoustic information when fused late.
    Foundation of ALFR; supported empirically but assumed a priori for architecture design.
  • domain assumption MUStARD full-set evaluation for training-free BiCAL and official CMMA test split / audio-available subset are fair for claimed comparisons.
    Section IV-A; authors note split heterogeneity in prior work.
invented entities (3)
  • Bidirectional Charge Calibration (BiCAL) continuum independent evidence
    purpose: Operationalize opposing prompt “charges” so directional verdict bias cancels and a 0–15 vote recovers pragmatic confidence.
    Methodological construct, not a physical entity; independent evidence is empirical Macro-F1 lift vs DZS and random-paraphrase control.
  • Acoustic Late-Fusion Rescue (ALFR) 27-D joint vector independent evidence
    purpose: Concatenate calibrated vote, 6 openSMILE z-features, and 20 Omni one-hot probes for a shallow rescue classifier.
    Engineering assembly; value shown by ablations and rescue of weak backbones.
  • Four multi-class Omni perception probes (intensity/style/tone-match/expressiveness) no independent evidence
    purpose: Avoid binary audio collapse by routing audio into multi-class perception tasks used as features.
    Prompted task definitions invented for this pipeline; no external psychometrics validating the exact option sets.

pith-pipeline@v1.1.0-grok45 · 32142 in / 3718 out tokens · 49561 ms · 2026-07-14T07:02:21.376918+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of CHARM: Charge Calibration and Acoustic Rescue for LLM-based Multimodal Sarcasm Detection." pith.science (2026). https://pith.science/paper/JDNTHOWH

@misc{pith2026260711102,
  author       = {Pith},
  title        = {Pith review of: CHARM: Charge Calibration and Acoustic Rescue for LLM-based Multimodal Sarcasm Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDNTHOWH}},
  note         = {Machine review of arXiv:2607.11102}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Sarcasm detection, the identification of discrepancies between literal and intended meaning, is a fundamental task in affective computing. However, zero-shot instruction-tuned Large Language Models (LLMs) systematically over-predict the positive (sarcastic) class across the entire capability spectrum, while the prosodic cues humans rely on remain underexploited and transfer unevenly across languages. We introduce CHARM (Charge Calibration and Acoustic Rescue for Multimodal Sarcasm Detection), a training-free framework that couples two modules. Bidirectional Charge Calibration (BiCAL) steers the LLM toward opposing sarcastic and literal verdicts along a symmetric axis of charged prompts; the induced directional biases cancel by construction, and a simple aggregation recovers an unbiased pragmatic signal. Acoustic Late-Fusion Rescue (ALFR) then fuses the calibrated votes with prosodic descriptors and LLM-generated auditory-perception probes through a shallow classifier, actively down-weighting saturated text votes in favour of acoustic evidence. Without fine-tuning any backbone, BiCAL attains the highest reported zero-shot text-only Macro-F1 of 0.787 on MUStARD, while ALFR lifts weak backbones by up to +0.382 Macro-F1 on CMMA. A Stouffer meta-analysis confirms statistical significance on MUStARD and CMMA (Z = 13.89 and Z = 34.64, respectively; p < 10^-43). Our analysis further uncovers a cross-cultural prosodic decoupling: low-level acoustics fail to transfer across languages, whereas high-level perceptual abstractions remain robust. Together, these components yield an explainable, cross-lingual multimodal detector.

Figures

Figures reproduced from arXiv: 2607.11102 by Bj\"orn W. Schuller, Qiyang Sun, Xi Shao, Yi Chang, Yupei Li, Zixing Zhang.

Figure 1
Figure 1. Figure 1: Architecture of the CHARM framework. Mode A (text-only BICAL): Five symmetric charge templates span from −2 to +2. These configurations elicit binary sarcasm judgements from a frozen LLM via stochastic decoding (T =0.5). Aggregating predictions across three random seeds yields fifteen cell-level votes. This symmetric layout cancels prior-induced token bias. The pipeline outputs a calibrated pragmatic confi… view at source ↗
Figure 2
Figure 2. Figure 2: The Symmetrical Bidirectional Charge Prompting Continuum within the BiCAL Module. A shared [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Verbatim runtime templates for the four Multi-Class Omni Perception Probes. Each probe isolates the exact prompt template executed in the production [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Random Forest impurity-based feature importance [43] allocated to [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 13 linked inside Pith

  1. [1]

    Sarcasm de- tection in the era of ai: A systematic review of techniques, datasets, performance, and future outlook,

    F. Santosa, E. Utami, K. Kusrini, and F. W. Wibowo, “Sarcasm de- tection in the era of ai: A systematic review of techniques, datasets, performance, and future outlook,” in2025 International Conference on Computer, Control, Informatics and its Applications (IC3INA). IEEE, 2025, pp. 19–24

    2025

  2. [2]

    R. W. Picard,Affective Computing. Cambridge, MA: MIT Press, 1997

    1997

  3. [3]

    Towards friendly ai: A comprehensive review and new perspectives on human-ai alignment,

    Q. Sun, Y . Li, E. Alturki, S. M. K. Murthy, and B. W. Schuller, “Towards friendly ai: A comprehensive review and new perspectives on human-ai alignment,”AI and Ethics, vol. 6, no. 2, p. 193, 2026

    2026

  4. [4]

    The muse 2024 multimodal sentiment analysis challenge: Social perception and humor recognition,

    S. Amiriparian, L. Christ, A. Kathan, M. Gerczuk, N. M ¨uller, S. Klug, L. Stappen, A. K ¨onig, E. Cambria, B. W. Schulleret al., “The muse 2024 multimodal sentiment analysis challenge: Social perception and humor recognition,” inProceedings of the 5th on Multimodal Sentiment Analysis Challenge and Workshop: Social Perception and Humor, 2024, pp. 1–9

    2024

  5. [5]

    Gatedxlstm: a multimodal affective computing approach for emotion recognition in conversations,

    Y . Li, Q. Sun, S. M. K. Murthy, E. Alturki, and B. W. Schuller, “Gatedxlstm: a multimodal affective computing approach for emotion recognition in conversations,”arXiv preprint arXiv:2503.20919, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Spoken in jest, detected in earnest: A systematic review of sarcasm recognition-multimodal fusion, challenges, and future prospects,

    X. Gao, S. Nayak, and M. Coler, “Spoken in jest, detected in earnest: A systematic review of sarcasm recognition-multimodal fusion, challenges, and future prospects,”IEEE Transactions on Affective Computing, 2025

    2025

  7. [7]

    Enhancing sarcasm detection in sentiment analysis for cyberspace safety using advanced deep learning techniques,

    R. Dhumpati, A. Sasi, S. J. Basha, S. Vadhri, P. Saroja, S. Alluri, and R. Vatambeti, “Enhancing sarcasm detection in sentiment analysis for cyberspace safety using advanced deep learning techniques,”Scientific Reports, vol. 15, no. 1, p. 22681, 2025

    2025

  8. [8]

    Is llama 3 good at sarcasm detection? a comprehensive study,

    Z. Mai, J. Zhang, Z. Xu, and Z. Xiao, “Is llama 3 good at sarcasm detection? a comprehensive study,” inProceedings of the 2024 7th In- ternational Conference on Machine Learning and Machine Intelligence (MLMI), 2024, pp. 141–145

    2024

  9. [9]

    Refashioning emotion recognition modeling: the advent of generalized large models,

    Z. Zhang, L. Peng, T. Pang, J. Han, H. Zhao, and B. W. Schuller, “Refashioning emotion recognition modeling: the advent of generalized large models,”IEEE Transactions on Computational Social Systems, vol. 11, no. 5, pp. 6690–6704, 2024

    2024

  10. [10]

    Affective computing in the era of large language models: A survey from the nlp perspective,

    Y . Zhang, X. Yang, X. Xu, Z. Gao, Y . Huang, S. Mu, S. Feng, D. Wang, Y . Zhang, K. Songet al., “Affective computing in the era of large language models: A survey from the nlp perspective,”Knowledge-Based Systems, p. 115411, 2026

    2026

  11. [11]

    Discovering language model behaviors with model-written evaluations,

    E. Perez, S. Ringer, K. Lukosiute, K. Nguyen, E. Chen, S. Heiner, C. Pettit, C. Olsson, S. Kundu, S. Kadavathet al., “Discovering language model behaviors with model-written evaluations,” inFindings of the association for computational linguistics: ACL 2023, 2023, pp. 13 387– 13 434

    2023

  12. [12]

    Towards understanding sycophancy in language models,

    M. Sharma, M. Tong, T. Korbak, D. Duvenaud, A. Askell, S. Bowman, E. Durmus, Z. Hatfield-Dodds, S. Johnston, S. Kravecet al., “Towards understanding sycophancy in language models,” inInternational Con- ference on Learning Representations, vol. 2024, 2024, pp. 110–144

    2024

  13. [13]

    Sarcasmbench: Towards evaluating large language models on sarcasm understanding,

    Y . Zhang, C. Zou, Z. Lian, P. Tiwari, and J. Qin, “Sarcasmbench: Towards evaluating large language models on sarcasm understanding,” IEEE Transactions on Affective Computing, vol. 16, no. 4, pp. 2560– 2578, 2025

    2025

  14. [14]

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

    B. Tang, B. Lin, H. Yan, and S. Li, “Leveraging generative large language models with visual instruction and demonstration retrieval for multimodal sarcasm detection,” inProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2024, pp. 1732–1742

    2024

  15. [15]

    Calibrate before use: Improving few-shot performance of language models,

    Z. Zhao, E. Wallace, S. Feng, D. Klein, and S. Singh, “Calibrate before use: Improving few-shot performance of language models,” in International conference on machine learning. Pmlr, 2021, pp. 12 697– 12 706

    2021

  16. [16]

    Language models don’t always say what they think: Unfaithful explanations in chain- of-thought prompting,

    M. Turpin, J. Michael, E. Perez, and S. Bowman, “Language models don’t always say what they think: Unfaithful explanations in chain- of-thought prompting,”Advances in Neural Information Processing Systems, vol. 36, pp. 74 952–74 965, 2023

    2023

  17. [17]

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

    B. Yao, Y . Zhang, Q. Li, and J. Qin, “Is sarcasm detection a step-by- step reasoning process in large language models?” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 24, 2025, pp. 25 651–25 659

    2025

  18. [18]

    Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Towards holistic evaluation of large audio-language models: A comprehensive survey,

    C.-K. Yang, N. S. Ho, and H.-y. Lee, “Towards holistic evaluation of large audio-language models: A comprehensive survey,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025, pp. 10 155–10 181

    2025

  20. [20]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  21. [21]

    Qwen2. 5-omni technical report, 2025,

    J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y . Fan, K. Danget al., “Qwen2. 5-omni technical report, 2025,”URL https://arxiv. org/abs/2503.20215, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    When language overrules: Revealing text dominance in multimodal large language models,

    H. Wu, M. Tang, X. Zheng, and H. Jiang, “When language overrules: Revealing text dominance in multimodal large language models,”arXiv preprint arXiv:2508.10552, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Rethinking explainability in the era of multimodal ai,

    C. Agarwal, “Rethinking explainability in the era of multimodal ai,” arXiv preprint arXiv:2506.13060, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Do audio llms really listen, or just transcribe? measuring lexical vs. acoustic emotion cues reliance,

    J. Chen, Z. Guo, J. Chun, P. Wang, A. Perrault, and M. Elsner, “Do audio llms really listen, or just transcribe? measuring lexical vs. acoustic emotion cues reliance,” inProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 2026, pp. 5848–5877

    2026

  25. [25]

    Explainable artificial in- telligence for medical applications: A review,

    Q. Sun, A. Akman, and B. W. Schuller, “Explainable artificial in- telligence for medical applications: A review,”ACM Transactions on Computing for Healthcare, vol. 6, no. 2, pp. 1–31, 2025

    2025

  26. [26]

    Improving audio explanations using audio language models,

    A. Akman, Q. Sun, and B. W. Schuller, “Improving audio explanations using audio language models,”IEEE Signal Processing Letters, vol. 32, pp. 741–745, 2025

    2025

  27. [27]

    Opensmile: the munich versatile and fast open-source audio feature extractor,

    F. Eyben, M. W ¨ollmer, and B. Schuller, “Opensmile: the munich versatile and fast open-source audio feature extractor,” inProceedings of the 18th ACM international conference on Multimedia, 2010, pp. 1459–1462

    2010

  28. [28]

    Towards multimodal sarcasm detection (an obviously perfect paper),

    S. Castro, D. Hazarika, V . P ´erez-Rosas, R. Zimmermann, R. Mihalcea, and S. Poria, “Towards multimodal sarcasm detection (an obviously perfect paper),” inProceedings of the 57th annual meeting of the association for computational linguistics, 2019, pp. 4619–4629

    2019

  29. [29]

    Cmma: benchmarking multi-affection detection in chinese multi-modal conversations,

    Y . Zhang, Y . Yu, Q. Guo, B. Wang, D. Zhao, S. Uprety, D. Song, Q. Li, and J. Qin, “Cmma: benchmarking multi-affection detection in chinese multi-modal conversations,”Advances in Neural Information Processing Systems, vol. 36, pp. 18 794–18 805, 2023

    2023

  30. [30]

    Leveraging large language models for sarcastic speech annotation in sarcasm detection,

    Z. Li, Y . Zhang, X. Gao, S. Nayak, and M. Coler, “Leveraging large language models for sarcastic speech annotation in sarcasm detection,” arXiv preprint arXiv:2506.00955, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Empathy omni: Enabling empathetic speech response generation through large language models,

    H. Wang, G. Zhang, J. Chen, J. Li, Y . Wang, and Y . Guo, “Empathy omni: Enabling empathetic speech response generation through large language models,”arXiv preprint arXiv:2508.18655, 2025

    arXiv 2025

  32. [32]

    Be- yond silent letters: Amplifying llms in emotion recognition with vocal nuances,

    Z. Wu, Z. Gong, L. Ai, P. Shi, K. Donbekci, and J. Hirschberg, “Be- yond silent letters: Amplifying llms in emotion recognition with vocal nuances,” inFindings of the Association for Computational Linguistics: NAACL 2025, 2025, pp. 2202–2218

    2025

  33. [33]

    Llms will always hallucinate, and we need to live with this,

    S. Banerjee, A. Agarwal, and S. Singla, “Llms will always hallucinate, and we need to live with this,” inIntelligent Systems Conference. Springer, 2025, pp. 624–648

    2025

  34. [34]

    Amused: An attentive deep neural network for multimodal sarcasm detection incorporating bi-modal data augmentation,

    X. Gao, S. Bansal, K. Gowda, Z. Li, S. Nayak, N. Kumar, and M. Coler, “Amused: An attentive deep neural network for multimodal sarcasm detection incorporating bi-modal data augmentation,”arXiv preprint arXiv:2412.10103, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Grainy-xai: A domain- wise granular analysis of sarcastic text using explainable-ai techniques

    J. R. Saini, S. Vaidya, I. Dhulekar, and K. Pal, “Grainy-xai: A domain- wise granular analysis of sarcastic text using explainable-ai techniques.” International Journal of Intelligent Engineering & Systems, vol. 17, no. 4, 2024

    2024

  36. [36]

    Qwen2.5 technical report,

    Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. Ren, X. Ren, Y . Fan, Y . Su, Y . Zhang, Y . Wan, Y . Liu, Z. Cui, Z. Zhang, ...

  37. [37]

    Available: https://arxiv.org/abs/2412.15115

    [Online]. Available: https://arxiv.org/abs/2412.15115

    Pith/arXiv arXiv

  38. [38]

    The llama 3 herd of models,

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Gemma 2: Improving open language models at a practical size,

    G. Team, M. Riviere, S. Pathak, P. G. Sessa, C. Hardin, S. Bhupatiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ram ´eet al., “Gemma 2: Improving open language models at a practical size,”arXiv preprint arXiv:2408.00118, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Mixtral of experts,

    A. Q. Jiang, A. Sablayrolles, A. Roux, A. Mensch, B. Savary, C. Bam- ford, D. S. Chaplot, D. d. l. Casas, E. B. Hanna, F. Bressandet al., “Mixtral of experts,”arXiv preprint arXiv:2401.04088, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 18

    Pith/arXiv arXiv 2024

  41. [41]

    Deepseek-v3 technical report,

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Openai gpt-5 system card,

    A. Singh, A. Fry, A. Perelman, A. Tart, A. Ganesh, A. El-Kishky, A. McLaughlin, A. Low, A. Ostrow, A. Ananthramet al., “Openai gpt-5 system card,”arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    D. W. Hosmer Jr, S. Lemeshow, and R. X. Sturdivant,Applied logistic regression. John Wiley & Sons, 2013

    2013

  44. [44]

    Random forests,

    L. Breiman, “Random forests,”Machine learning, vol. 45, no. 1, pp. 5–32, 2001

    2001

  45. [45]

    Greedy function approximation: a gradient boosting machine,

    J. H. Friedman, “Greedy function approximation: a gradient boosting machine,”Annals of statistics, pp. 1189–1232, 2001

    2001

  46. [46]

    A decision-theoretic generalization of on-line learning and an application to boosting,

    Y . Freund and R. E. Schapire, “A decision-theoretic generalization of on-line learning and an application to boosting,”Journal of computer and system sciences, vol. 55, no. 1, pp. 119–139, 1997

    1997

  47. [47]

    Support vector regression machines,

    H. Drucker, C. J. Burges, L. Kaufman, A. Smola, and V . Vapnik, “Support vector regression machines,”Advances in neural information processing systems, vol. 9, 1996

    1996

  48. [48]

    Multilayer perceptron and neural networks,

    M.-C. Popescu, V . E. Balas, L. Perescu-Popescu, and N. Mastorakis, “Multilayer perceptron and neural networks,”WSEAS transactions on circuits and systems, vol. 8, no. 7, pp. 579–588, 2009

    2009

  49. [49]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” inProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 2016, pp. 785–794

    2016

  50. [50]

    Lightgbm: A highly efficient gradient boosting decision tree,

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.- Y . Liu, “Lightgbm: A highly efficient gradient boosting decision tree,” Advances in neural information processing systems, vol. 30, 2017

    2017

  51. [51]

    Stacked generalization,

    D. H. Wolpert, “Stacked generalization,”Neural networks, vol. 5, no. 2, pp. 241–259, 1992

    1992

  52. [52]

    On combining classifiers,

    J. Kittler, M. Hatef, R. P. Duin, and J. Matas, “On combining classifiers,” IEEE transactions on pattern analysis and machine intelligence, vol. 20, no. 3, pp. 226–239, 1998

    1998

  53. [53]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language mod- els are few-shot learners,”Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  54. [54]

    Combining probability from independent tests: the weighted z-method is superior to fisher’s approach,

    M. C. Whitlock, “Combining probability from independent tests: the weighted z-method is superior to fisher’s approach,”Journal of evolu- tionary biology, vol. 18, no. 5, pp. 1368–1373, 2005

    2005

  55. [55]

    When to use the b onferroni correction,

    R. A. Armstrong, “When to use the b onferroni correction,”Ophthalmic and physiological optics, vol. 34, no. 5, pp. 502–508, 2014

    2014

  56. [56]

    A deep learning based contrastive atten- tion model for multi-modal sarcasm detection,

    V . Sukhavasi and V . Dondeti, “A deep learning based contrastive atten- tion model for multi-modal sarcasm detection,” in2025 International Conference on Computing Technologies (ICOCT). IEEE, 2025, pp. 1–6

    2025

  57. [57]

    Breakthrough from nuance and inconsistency: Enhancing multimodal sarcasm detection with context-aware self-attention fusion and word weight calculation

    H. Xue, L. Xu, Y . Tong, R. Li, J. Lin, and D. Jiang, “Breakthrough from nuance and inconsistency: Enhancing multimodal sarcasm detection with context-aware self-attention fusion and word weight calculation.” in Proceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024...

    2024

  58. [58]

    How capable are state-of-the-art language models to cope with sarcasm?

    A.-C. B ˘aroiu and S ¸. Tr˘aus ¸an-Matu, “How capable are state-of-the-art language models to cope with sarcasm?” in2023 24th International Conference on Control Systems and Computer Science (CSCS). IEEE, 2023, pp. 399–402

    2023

  59. [59]

    Pragmatic metacognitive prompting improves llm performance on sarcasm detec- tion,

    J. Lee, W. Fong, A. Le, S. Shah, K. Han, and K. Zhu, “Pragmatic metacognitive prompting improves llm performance on sarcasm detec- tion,” inProceedings of the 1st Workshop on Computational Humor (CHum), 2025, pp. 63–70

    2025

  60. [60]

    Masd: A multi-agent sarcasm detection framework with chain-of-thought,

    Y . Zheng and L. Ma, “Masd: A multi-agent sarcasm detection framework with chain-of-thought,” in2025 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2025, pp. 1087–1092

    2025

  61. [61]

    Sarc7: Evaluating sarcasm detection and generation with seven types and emotion-informed techniques,

    L. Xiong, R. Gao, and A. Jeong, “Sarc7: Evaluating sarcasm detection and generation with seven types and emotion-informed techniques,” in Proceedings of the 9th Widening NLP Workshop, 2025, pp. 157–166

    2025

  62. [62]

    Acoustic markers of sarcasm in cantonese and english,

    H. S. Cheang and M. D. Pell, “Acoustic markers of sarcasm in cantonese and english,”The Journal of the Acoustical Society of America, vol. 126, no. 3, pp. 1394–1405, 2009

    2009

  63. [63]

    The role of voice quality in mandarin sarcastic speech: An acoustic and electroglottographic study,

    S. Li, W. Gu, L. Liu, and P. Tang, “The role of voice quality in mandarin sarcastic speech: An acoustic and electroglottographic study,”Journal of Speech, Language, and Hearing Research, vol. 63, no. 8, pp. 2578– 2588, 2020

    2020

  64. [64]

    The pragmatics of verbal irony: Echo or pretence?

    D. Wilson, “The pragmatics of verbal irony: Echo or pretence?”Lingua, vol. 116, no. 10, pp. 1722–1743, 2006

    2006