Pith. sign in

REVIEW 4 major objections 4 minor 48 references

Evaluating Vision-Language Models for Emotion Recognition

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

Pith's one-line read Zero-shot vision-language models are inept at recognizing emotions evoked by images, and their benchmark scores swing with prompt label order, persona, and label availability, while many residual errors trace to noisy dataset labels.

desk verdict A genuinely useful multi-dataset VLM emotion benchmark with a good error taxonomy, but the headline claim is undercut by unreported JSON parse failures and single-run evaluations. read the letter →

arxiv 2502.05660 v1 pith:DIB27L4M submitted 2025-02-08 cs.CV cs.CL

classification cs.CVcs.CL
keywords evokedemotionrecognitionvision-languagemodelszero-shotevaluationpromptrobustnessaffectivecomputingbenchmarkhuman
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 sets out to show that current large vision-language models, evaluated zero-shot, are poor at predicting the emotion an image evokes in a viewer. To do so it builds EVE, an Evoked Emotion benchmark assembled from five existing image-emotion datasets, with the two largest downsampled to hard subsets, and evaluates seven VLMs under eight prompt variations. The results show that performance is strongly shaped by prompt format: label order, the presence or absence of target labels, and an adopted emotional persona each shift scores and sentiment bias by large margins. A blinded human study of roughly five hundred error cases indicates that the broadest errors are genuine model failures, while the finest-grained mismatches are often judged by humans as better than, or equal to, the dataset's own label. If the conclusion holds, affective benchmarks should report prompt-dependent score ranges and emotion datasets need multi-label or explanation-based labels before VLM fine-tuning can be meaningfully assessed.

What carries the argument

The load-bearing machinery is the EVE benchmark combined with a three-part error-analysis framework. EVE unites EmoSet-Hard, FI-Hard, Abstract, ArtPhoto, and Emotion6, where the hard subsets are created by fine-tuning a ViT and keeping its misclassified samples together with correct samples predicted at softmax confidence below an empirically chosen 0.8. The analysis framework defines a sentiment bias pair—positive bias $p_p = p(c \in S_P \mid l \in S_N)$ and negative bias $p_n = p(c \in S_N \mid l \in S_P)$—and categorizes every mistake into three levels: wrong sentiment, right sentiment but wrong arousal, and right sentiment and arousal but wrong fine-grained class. A blinded human-agreement study of about 500 error samples is then attached to this taxonomy. Together these pieces let the authors attribute observed F1 gaps either to model capability or to unreliable labels.

What would settle it

Re-run the EVE evaluation with hard subsets selected by inter-human disagreement or by a second independently trained ViT, instead of by the original ViT's softmax confidence; if the low F1 scores and the label-order and persona sensitivities shrink or reverse under that filter, the central ineptness claim is an artifact of the difficulty proxy. A complementary decisive observation would be a zero-shot VLM whose weighted F1 on EVE exceeds fine-tuned task-specific models across all eight prompt variants.

Watch

Extended reading notes

Core claim

The central finding is that at the current state, VLMs are inept at zero-shot evoked emotion recognition: weighted F1 scores on EVE range from roughly 0.08 on Abstract images to about 0.64 on Emotion6, and no model matches task-specific fine-tuned systems overall. The paper further shows that this poor performance is not a fixed model trait. Listing negative emotion labels first lowers accuracy for open models, removing labels entirely hurts every model even when predictions are matched by semantic similarity, and prompting an optimistic or pessimistic persona degrades accuracy while driving predictions toward one sentiment. Reasoning prompts help only selectively: caption-then-reason benefits LLaVA-family models, whereas three-step contextual reasoning helps mainly GPT4-o. The error analysis then separates model failure from benchmark failure: sentiment-level errors are confirmed as model errors by human raters, but for the most fine-grained error category the raters prefer the model's predicted label or both labels over the dataset ground truth, implying that a substantial share of the apparent deficiency is caused by noisy or subjective ground truth.

Load-bearing premise

The load-bearing premise is that the softmax confidence of a fine-tuned ViT, with a cutoff empirically set to 0.8, is a valid measure of how hard an image is for humans or VLMs; if that proxy is wrong, the hard subsets of EVE are arbitrary and the reported F1 scores do not measure evoked emotion recognition in general.

Editorial extensions

If this is right

  • On EVE, zero-shot weighted F1 scores stay in a low band—roughly 0.08 to 0.64—so no current VLM is reliable for evoked emotion recognition without adaptation.
  • Benchmark rankings are prompt-relative: reordering or removing target labels and adding an emotional persona each change scores and flip sentiment bias, so a single F1 number cannot summarize a model's affective ability.
  • Models show a default positive bias when labels are supplied and a default negative bias in open-vocabulary prediction, meaning the bias is an artifact of prompt anchoring rather than a stable model personality.
  • Reasoning prompts do not fix the deficiency; caption-based reasoning gives modest gains to open LLaVA-family models, and contextual reasoning helps mostly GPT4-o, while the remaining models struggle with the multi-step format.
  • For the finest-grained error category, human raters agree with the model's prediction or with both labels more often than with the dataset ground truth, so a portion of the low scores reflects noisy labels rather than VLM deficit.

Reading between the lines

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

  • The ViT-confidence difficulty filter defines hard samples as those that puzzle one particular fine-tuned ViT; replacing it with a human-disagreement-based filter could change which images enter EVE and therefore alter the reported F1 gaps, so the benchmark's difficulty is partly an artifact of its own proxy.
  • A testable extension of the prompt-sensitivity result is to require affective models to be invariant across the eight prompt variants before being ranked; models that are non-invariant should be reported with score distributions rather than point estimates.
  • The especially poor performance on abstract paintings suggests that VLM emotion judgments may be carried largely by object and scene recognition rather than affective reasoning; comparing representational versus abstract image pairs would isolate that mechanism.
  • Because the human raters were a small group of graduate-student volunteers in North America, the claim that fine-grained errors are largely noisy ground truth could be stress-tested by repeating the annotation with a larger, culturally diverse rater pool, since emotion perception varies across cultures.
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 / 4 minor

Summary. The paper introduces EVE, a benchmark for evoked emotion recognition built from five existing datasets, with EmoSet and FI downsampled to 'hard' subsets using the softmax confidence of a fine-tuned ViT. The authors evaluate seven open-source VLMs and GPT-4o in zero-shot simple classification and in prompt-robustness conditions (label-order shuffling, open-vocabulary prediction, positive/negative personas, and three reasoning schemes), reporting weighted F1 scores and sentiment-bias measures. They also define a three-level error taxonomy and run a human annotation study on about 500 error samples to attribute errors to model limitations versus ground-truth noise and task subjectivity. The central claim is that current VLMs are inept at evoked emotion recognition, are strongly sensitive to prompt format, and that many fine-grained errors reflect unreliable dataset labels rather than model inability.

Significance. If the findings withstand scrutiny, the paper would provide a useful first benchmark and a cautionary reference for affective VLM research, with the public release of code/data, detailed prompt transcripts, an explicit error taxonomy, and an attempt to connect machine errors to human perception as genuine strengths. The main contribution is diagnostic rather than methodological: it identifies prompt-format sensitivity and label noise as first-order issues in VLM emotion evaluation. However, the quantitative conclusions are currently weakened by unmeasured format-compliance failures, single-run evaluations with no variance estimates, an empirically chosen difficulty filter, and a small in-group human study. These issues do not necessarily overturn the qualitative direction of the findings, but they must be addressed before the specific F1 values and prompt-sensitivity comparisons can be treated as reliable.

major comments (4)
  1. [Sec. 4; Figs. 12-20] The evaluation pipeline is described in Section 4 as 'responses are parsed and string-matched with the ground truth class labels,' but the paper never reports the fraction of model outputs that conform to the required JSON schema, the parser's success rate, or how malformed outputs are scored (discarded versus counted as errors). Since every prompt demands strict JSON output and the reasoning prompts demand multi-key JSON, any non-trivial format-failure rate would conflate emotion recognition with instruction following. This threat is uneven across conditions because the persona and reasoning prompts are more complex than the simple-classification prompt, so format failures could masquerade as prompt-sensitivity effects. Please report parse-success rates per model and per condition, and re-analyze the results after excluding or explicitly accounting for malformed responses.
  2. [Sec. 3; Sec. A.1] The benchmark's difficulty filter uses the softmax confidence of a fine-tuned ViT with a threshold 'chosen empirically to be 0.8,' determined from the prediction-probability distribution of the same ViT. This threshold is a free parameter that selects which images enter EVE, and no independent evidence establishes that ViT confidence tracks human-perceived or VLM-perceived difficulty. The aggregate F1 scores therefore describe a self-selected filtered subset rather than evoked emotion recognition over the original datasets. Please provide a principled selection criterion, ablate the threshold, or clearly restrict the claims to the filtered subset so that the benchmark's scope is transparent.
  3. [Sec. 7; Sec. B.4.4] The human study uses roughly 500 error samples annotated by 'graduate student volunteers from the authors' team,' with no reported number of annotators, no inter-annotator agreement metric, and no detail on how disagreements were resolved. The RQ3 conclusion that EC II and EC III errors largely reflect subjective or noisy ground truth rests on this study, so the absence of reliability evidence is load-bearing for the paper's central error-analysis claim. Please use multiple independent annotators, report agreement statistics, and verify that the agreement percentages in Fig. 7 are robust across annotators.
  4. [Sec. B.1; Sec. 6] Appendix B.1 states that all results come from single runs, but the main text reports only point estimates without variance, confidence intervals, or significance tests. The robustness claims in Section 6, such as the label-order effects in Fig. 2 and the persona effects in Fig. 5, rely on differences that could be within run-to-run variability of stochastic model decoding. Please provide multiple runs with variance estimates, or justify deterministic decoding and report the sampling settings, so that the prompt-sensitivity conclusions can be assessed quantitatively.
minor comments (4)
  1. [Sec. 9] The Limitations section states that 'the current evaluation also includes only few-shot performances of the models,' but the experiments are zero-shot; this appears to be a typo and should be corrected.
  2. [Sec. 5] The sentence 'Models from the the LLaV A family' contains a duplicated 'the,' and later in the same paragraph 'the largest different between' should be 'the largest difference between.'
  3. [Figs. 13-20] The prompt figures render the prompts as continuous strings without preserving line breaks, which makes the exact instructions, especially the reasoning prompts, harder to read and reproduce; providing a plain-text listing in the appendix would improve reproducibility.
  4. [Sec. 1] The phrase 'atleast partially present' in the introduction should be 'at least partially present'; there are also several other minor spacing errors in the prompt transcriptions in Appendix B.3 that should be normalized.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the evaluation is self-contained; the single fitted threshold selects data without entering the scoring equations, and self-citations appear only as background and recommendations.

full rationale

The paper performs an empirical evaluation rather than a derivation, so the standard circularity patterns do not apply. Benchmark construction is the only place a value is fitted: a fine-tuned ViT's prediction probabilities are used with a threshold 'chosen empirically to be 0.8' to build EmoSet-Hard and FI-Hard (Section 3, A.1). That threshold selects which samples enter the benchmark but never appears in the weighted-F1 scoring equations, and the central 'VLMs are inept' conclusion is additionally supported on the unfiltered datasets (Emotion6, Abstract, ArtPhoto, Table 1), so no reported score is equal to an input by construction. The prompt-robustness findings (label order, persona, open vocabulary, reasoning) are direct measurements of model outputs; the open-vocabulary assignment relies on SBERT, an external model with no parameters fitted in this paper, and the fine-grained-ness criterion simmax is defined only from the class-label set. The error analysis (RQ3) derives its conclusions from blinded human annotations of about 500 samples; the claim that EC-III disagreements reflect noisy ground truth is an empirical tally (annotators preferred model labels over dataset labels for those cases), not a definitional equivalence. Self-citations (Wang et al. 2023a; Kim et al. 2018; Lu et al. 2017) are used only as background motivation and as future-data recommendations; they are not invoked to justify any measured quantity, to import a uniqueness theorem, or to adopt an ansatz. The unreported JSON parse-success rate and the reliance on the authors' own graduate students as annotators are internal-validity concerns (correctness risk), not circularity: neither makes a reported score reduce to a fitted quantity. Hence no load-bearing circular step; score 1.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central evaluation rests on the benchmark being a valid measure of evoked emotion recognition, which in turn rests on the ViT-confidence difficulty filter, the reliability of existing labels, the neutrality of the human-perspective prompt, and the representativeness of the authors' human annotations. These are domain assumptions rather than demonstrated facts. No new entities are introduced.

free parameters (2)
  • difficulty threshold for hard subset selection = 0.8
    Chosen empirically from the distribution of ViT prediction probabilities on EmoSet and FI to define intermediate difficulty samples (Section 3, A.1).
  • target subset size for downsampled datasets = about 2900 samples each for EmoSet and FI
    Arbitrary, cost-driven downsample target; affects statistical power and benchmark composition (Section 3).
assumptions (5)
  • domain assumption Softmax confidence of a fine-tuned ViT is a valid measure of emotion-recognition difficulty.
    Underlies the construction of EmoSet-Hard and FI-Hard in Section 3 and A.1; if false, the benchmark is not actually hard in a meaningful sense.
  • domain assumption Existing dataset labels are treated as ground truth for F1.
    All main experiments in Section 5 score against these labels, while Section 7 later argues many are noisy.
  • ad hoc to paper The instruction to imagine being a human does not change emotion recognition behavior beyond bypassing moderation.
    Appendix B.3 states this line was added to obtain responses; its effect on predictions is not measured.
  • domain assumption SBERT similarity is a valid mapping from free-form emotion words to the fixed class set.
    Used for open-vocabulary evaluation in Section 6.2; different embeddings would give different assignments.
  • domain assumption Annotations by the authors' graduate-student volunteers are a reliable measure of human emotion agreement.
    Section 7 and B.4.4 base the noisy-ground-truth claim on these annotations; no inter-annotator agreement is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Vision-Language Models for Emotion Recognition." pith.science (2026). https://pith.science/paper/DIB27L4M

@misc{pith2026250205660,
  author       = {Pith},
  title        = {Pith review of: Evaluating Vision-Language Models for Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIB27L4M}},
  note         = {Machine review of arXiv:2502.05660}
}
read the original abstract

Large Vision-Language Models (VLMs) have achieved unprecedented success in several objective multimodal reasoning tasks. However, to further enhance their capabilities of empathetic and effective communication with humans, improving how VLMs process and understand emotions is crucial. Despite significant research attention on improving affective understanding, there is a lack of detailed evaluations of VLMs for emotion-related tasks, which can potentially help inform downstream fine-tuning efforts. In this work, we present the first comprehensive evaluation of VLMs for recognizing evoked emotions from images. We create a benchmark for the task of evoked emotion recognition and study the performance of VLMs for this task, from perspectives of correctness and robustness. Through several experiments, we demonstrate important factors that emotion recognition performance depends on, and also characterize the various errors made by VLMs in the process. Finally, we pinpoint potential causes for errors through a human evaluation study. We use our experimental results to inform recommendations for the future of emotion research in the context of VLMs.

Figures

Figures reproduced from arXiv: 2502.05660 by the authors.

Figure 1
Figure 1. Positive and Negative Bias demonstrated by [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. (a): The weighted F1 score for each model, averaged across datasets. The different bars represent the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The weighted F1 score with and without precise target labels in the prompts. The numbers in brown represent the percentage of fine-grained predic￾tions made. Qwen-VL LLaVA LLaVA-Next GPT4-o Positive Negative 0.24 0.16 0.24 0.51 0.4 0.59 0.41 0.12 0.25 0.50 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: Sentiment bias for responses generated with [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Fig. (a): Weighted F1 score for each model, averaged across all datasets considered. The score drops [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Weighted F1 scores, averaged across all datasets considered, for different prompting mecha￾nisms explored. tion for its emotion prediction simultaneously. The second mechanism uses three steps of contextual reasoning prior to prediction. The first two steps in￾volve at…
Figure 7
Figure 7. Figure 7: Human agreement with model predictions (blue), ground truth from dataset (orange), both labels (green), [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The distribution of different emotion classes in the final evaluation sets considered. The numbers of [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: The distribution of different emotion classes in the final evaluation sets considered, grouped according to [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Examples from the created EmoSet-Hard dataset. For Contentment and Excitement, no instances are found that are predicted correctly with a probability less than 0.8. For all other categories, the two leftmost examples describe instances that are correctly predicted, bu…
Figure 11
Figure 11. Figure 11: Examples from the created FI-Hard dataset. Similar to EmoSet-Hard, for Contentment and Sadness, no instances are found that are predicted correctly with a probability less than 0.8. For all other categories, the two leftmost examples describe instances that are correc…
Figure 12
Figure 12. Figure 12: The prompt Simple Multimodal Classification [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: The prompt for shuffled order of emotions with positive emotions first. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: The prompt for shuffled order of emotions with negative emotions first. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: The prompt for open-vocabulary emotion prediction. [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 16
Figure 16. Figure 16: The prompt for adopting positive persona. [PITH_FULL_IMAGE:figures/full_fig_p018_16.png]
Figure 17
Figure 17. Figure 17: The prompt for adopting negative persona. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]
Figure 18
Figure 18. Figure 18: The prompt Explanation-based Reasoning Imagine you are like a human, capable of feeling emotions, and an image is shown to you. This image may evoke positive or negative emotions. Your task is to answer what emotion is evoked in you when viewing the given image. Consi…
Figure 19
Figure 19. Figure 19: The prompt for Contextual Reasoning Imagine you are like a human, capable of feeling emotions, and an image is shown to you. This image may evoke positive or negative emotions. Your task is to answer what emotion is evoked in you when viewing the given image. Consider…
Figure 20
Figure 20. Figure 20: The prompt Caption-Based Reasoning [PITH_FULL_IMAGE:figures/full_fig_p019_20.png]
Figure 21
Figure 21. Figure 21: Average Difference between the most similar [PITH_FULL_IMAGE:figures/full_fig_p020_21.png]
Figure 22
Figure 22. Figure 22: The frequency of agreement with model predicted label (cream/beige), the dataset ground truth label (green), both labels (yellow), and neither label (orange), for each dataset, averaged across all models studied. ducted. To create the different error categories, as we…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 31 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. G pt-4 T echnical R eport. arXiv preprint arXiv:2303.08774

  2. [2]

    Md Rabiul Awal, Rui Cao, Roy Ka-Wei Lee, and Sandra Mitrovi \'c . 2021. A ngrybert: J oint L earning T arget A nd E motion F or H ate S peech D etection. In Pacific-Asia conference on knowledge discovery and data mining, pages 701--713. Springer

  3. [3]

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Q wen- VL : A F rontier L arge V ision- L anguage M odel with V ersatile A bilities. arXiv preprint arXiv:2308.12966

  4. [4]

    Plummer, and Kate Saenko

    Katherine Deng, Arijit Ray, Reuben Tan, Saadia Gabriel, Bryan A. Plummer, and Kate Saenko. 2023. http://arxiv.org/abs/2308.16741 S ocratis: A re L arge M ultimodal M odels E motionally A ware? arXiv preprint arXiv:2308.16741

  5. [5]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. A n I mage is W orth 16x16 W ords: T ransformers for I mage R ecognition at S cale. In International Conference on Learning Representations

  6. [6]

    Yasaman Etesam, \"O zge Nilay Yal c n, Chuxuan Zhang, and Angelica Lim. 2024. C ontextual E motion R ecognition U sing L arge V ision L anguage M odels. arXiv preprint arXiv:2405.08992

  7. [7]

    Yi Fung, Ruining Zhao, Jae Doo, Chenkai Sun, and Heng Ji. 2024. M assively M ulti-cultural K nowledge A cquisition & L m B enchmarking. arXiv preprint arXiv:2402.09369

  8. [8]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. M easuring M assive M ultitask L anguage U nderstanding. In International Conference on Learning Representations

Show all 48 references
  1. [9]

    Manh-Hung Hoang, Soo-Hyung Kim, Hyung-Jeong Yang, and Guee-Sang Lee. 2021. https://doi.org/10.1109/ACCESS.2021.3091169 C ontext-aware E motion R ecognition B ased O n V isual R elationship D etection . IEEE Access, 9:90465--90474

  2. [10]

    Jyoti and Seema Rao. 2016. https://doi.org/10.1145/2979779.2979832 A S urvey O n S entiment A nalysis A nd O pinion M ining . In Proceedings of the International Conference on Advances in Information Communication Technology & Computing, AICTC '16, New York, NY, USA. Associati...

  3. [11]

    Aditi Khandelwal, Utkarsh Agarwal, Kumar Tanmay, and Monojit Choudhury. 2024. D o M oral J udgment A nd R easoning C apability O f L lms C hange W ith L anguage? A S tudy U sing T he M ultilingual D efining I ssues T est. arXiv preprint arXiv:2402.02135

  4. [12]

    Hanjoo Kim, Xin Lu, Michael Costa, Baris Kandemir, Reginald B Adams Jr, Jia Li, James Z Wang, and Michelle G Newman. 2018. D evelopment A nd V alidation O f I mage S timuli F or E motion E licitation ( isee): A N ovel A ffective P ictorial S ystem W ith T est-retest R epeatabi...

  5. [13]

    Agata Ko akowska, Agnieszka Landowska, Mariusz Szwoch, Wioleta Szwoch, and Michal R Wrobel. 2014. E motion R ecognition A nd I ts A pplications. Human-computer systems interaction: Backgrounds and applications 3, pages 51--62

  6. [14]

    Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoon Sohn. 2019. C ontext-aware E motion R ecognition N etworks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

  7. [15]

    Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. 2023. L arge L anguage M odels U nderstand A nd C an B e E nhanced B y E motional S timuli. arXiv preprint arXiv:2307.11760

  8. [16]

    CHENG LI, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Xinyi Wang, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. 2024. T he G ood, T he B ad, A nd W hy: U nveiling E motions I n G enerative A i. In Forty-first International Conference on Machine Learning

  9. [17]

    Zaijing Li, Gongwei Chen, Rui Shao, Dongmei Jiang, and Liqiang Nie. 2024. http://arxiv.org/abs/2401.06836 E nhancing E motional G eneration C apability of L arge L anguage M odels via E motional C hain-of- T hought . arXiv preprint arXiv:2401.06836. ArXiv:2401.06836 [cs]

  10. [18]

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024 a . https://llava-vl.github.io/blog/2024-01-30-llava-next/ LLaVA-NeXT : I mproved R easoning, OCR , and W orld K nowledge

  11. [19]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024 b . V isual I nstruction T uning. The 38th Conference on Neural Information Processing Systems (NeurIPS), 36

  12. [20]

    Zhiwei Liu, Kailai Yang, Tianlin Zhang, Qianqian Xie, Zeping Yu, and Sophia Ananiadou. 2024 c . E mollms: A S eries O f E motional L arge L anguage M odels A nd A nnotation T ools F or C omprehensive A ffective A nalysis. arXiv preprint arXiv:2401.08508

  13. [21]

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. L earn T o E xplain: M ultimodal R easoning V ia T hought C hains F or S cience Q uestion A nswering. In The 36th Conference on Neural Information P...

  14. [22]

    Xin Lu, Reginald B Adams, Jia Li, Michelle G Newman, and James Z Wang. 2017. A n I nvestigation I nto T hree V isual C haracteristics O f C omplex S cenes T hat E voke H uman E motion. In 2017 Seventh International Conference on Affective Computing and Intelligent Interaction ...

  15. [23]

    Jana Machajdik and Allan Hanbury. 2010. A ffective I mage C lassification using F eatures I nspired by P sychology and A rt T heory. In Proceedings of the 18th ACM International Conference on Multimedia, pages 83--92

  16. [24]

    Joseph A Mikels, Barbara L Fredrickson, Gregory R Larkin, Casey M Lindberg, Sam J Maglio, and Patricia A Reuter-Lorenz. 2005. E motional C ategory D ata on I mages from the I nternational A ffective P icture S ystem. Behavior Research Methods, 37:626--630

  17. [25]

    Trisha Mittal, Pooja Guhan, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. 2020. E moticon: C ontext-aware M ultimodal E motion R ecognition U sing F rege's P rinciple. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  18. [26]

    Vinod Nair and Geoffrey E Hinton. 2010. R ectified L inear U nits I mprove R estricted B oltzmann M achines. In Proceedings of the 27th International Conference on Machine Learning, pages 807--814

  19. [27]

    Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew Gallagher. 2015. https://doi.org/10.1109/CVPR.2015.7298687 A M ixed B ag of E motions: M odel, P redict, and T ransfer E motion D istributions . In Proceedings of the IEEE/CVF International Conference on Computer Vision, ...

  20. [28]

    Ciaran Regan, Nanami Iwahashi, Shogo Tanaka, and Mizuki Oka. 2024. https://doi.org/10.48550/arXiv.2402.04232 C an G enerative A gents P redict E motion? arXiv preprint arXiv:2402.04232. ArXiv:2402.04232 [cs]

  21. [29]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

  22. [30]

    Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. 2021. I magenet-21k P retraining F or T he M asses. In The 35th Conference on Neural Information Processing Systems (NeurIPS)(Datasets and Benchmarks Track)

  23. [31]

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. 2015. I magenet L arge S cale V isual R ecognition C hallenge. International Journal of Computer Vision, 115:211--252

  24. [32]

    Cobo, Carlos De Las Heras-Pedrosa, José Ignacio Peláez, and Enrique Herrera-Viedma

    Pablo Sánchez-Núñez, Manuel J. Cobo, Carlos De Las Heras-Pedrosa, José Ignacio Peláez, and Enrique Herrera-Viedma. 2020. https://doi.org/10.1109/ACCESS.2020.3009482 O pinion M ining, S entiment A nalysis A nd E motion U nderstanding I n A dvertising: A B ibliometric A nalysis ...

  25. [33]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. L lama: O pen A nd E fficient F oundation L anguage M odels. arXiv preprint arXiv:2302.13971

  26. [34]

    Wang, Sicheng Zhao, Chenyan Wu, Reginald B

    James Z. Wang, Sicheng Zhao, Chenyan Wu, Reginald B. Adams, Michelle G. Newman, Tal Shafir, and Rachelle Tsachor. 2023 a . https://doi.org/10.1109/JPROC.2023.3273517 U nlocking the E motional W orld of V isual M edia: A n O verview of the S cience, R esearch, and I mpact of U ...

  27. [35]

    Xuena Wang, Xueting Li, Zi Yin, Yue Wu, and Jia Liu. 2023 b . E motional I ntelligence of L arge L anguage M odels. Journal of Pacific Rim Psychology, 17:18344909231213958

  28. [36]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. C hain-of-thought P rompting E licits R easoning I n L arge L anguage M odels. The 36th Conference on Neural Information Processing Systems (NeurIPS), 35:24824--24837

  29. [37]

    Matthias J Wieser, Elisabeth Klupp, Peter Weyers, Paul Pauli, David Weise, Daniel Zeller, Joseph Classen, and Andreas M \"u hlberger. 2012. R educed E arly V isual E motion D iscrimination A s A n I ndex O f D iminished E motion P rocessing I n P arkinson’s D isease?--evidence...

  30. [38]

    BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ili \'c , Daniel Hesslow, Roman Castagn \'e , Alexandra Sasha Luccioni, Fran c ois Yvon, et al. 2022. B loom: A 1 76b-parameter O pen-access M ultilingual L anguage M odel. arXiv preprint ...

  31. [39]

    Alexandros Xenos, Niki Maria Foteinopoulou, Ioanna Ntinou, Ioannis Patras, and Georgios Tzimiropoulos. 2024. V llms P rovide B etter C ontext F or E motion U nderstanding T hrough C ommon S ense R easoning. arXiv preprint arXiv:2404.07078

  32. [40]

    Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. 2024. E movit: R evolutionizing E motion I nsights W ith V isual I nstruction T uning. arXiv preprint arXiv:2404.16670

  33. [41]

    Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. 2022. M dan: M ulti-level D ependent A ttention N etwork F or V isual E motion A nalysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9479--9488

  34. [42]

    Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischinski, Danny Cohen-Or, and Hui Huang. 2023. E mo S et: A L arge- S cale V isual E motion D ataset with R ich A ttributes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20383--20394

  35. [43]

    Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. 2021. S timuli-aware V isual E motion A nalysis. IEEE Transactions on Image Processing, 30:7432--7445

  36. [44]

    Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. 2016. B uilding a L arge S cale D ataset for I mage E motion R ecognition: T he F ine P rint and the B enchmark. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30

  37. [45]

    Sicheng Zhao, Hongxun Yao, Yue Gao, Guiguang Ding, and Tat-Seng Chua. 2018. https://doi.org/10.1109/TAFFC.2016.2628787 P redicting P ersonalized I mage E motion P erceptions I n S ocial N etworks . IEEE Transactions on Affective Computing, 9(4):526--540

  38. [46]

    Caleb Ziems, Jane Yu, Yi-Chia Wang, Alon Halevy, and Diyi Yang. 2022. https://doi.org/10.18653/v1/2022.acl-long.261 The moral integrity corpus: A benchmark for ethical dialogue systems . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics...

  39. [47]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  40. [48]

    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...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.