Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Hazard or Anomaly? Evaluating VLMs for Understanding Dangers and Discrepancies

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read VLMs routinely mistake unusual scenes for dangerous ones, a study of six vision-language models shows.

desk verdict Useful four-class benchmark, but the conflation claim needs a confusion matrix before it can be believed. read the letter →

arxiv 2607.18325 v1 pith:DFSR5VHP submitted 2026-07-18 cs.CV cs.AIcs.RO

Murali Indukuri , Mohammad Eskandari , Sree Nitya Kollu , Stephanie Lukin , Cynthia Matuszek This is my paper
classification cs.CVcs.AIcs.RO
keywords vision-languagemodelshazarddetectionanomalysafetyreasoningfour-classclassificationover-reactionproblempromptengineeringsituationalawareness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that current evaluations of vision-language models (VLMs) for safety are too coarse: they ask models to judge scenes as simply safe or unsafe, which hides a systematic failure mode. The authors introduce a four-class taxonomy—Safe, Anomalous, Hazardous, and Anomalous-Hazardous—and evaluate six VLMs across several prompting strategies on a new 610-image benchmark. Their central finding is that VLMs frequently classify merely anomalous scenes as hazardous, revealing an over-reliance on contextual irregularity as a proxy for danger. The paper also shows that explicitly separating anomaly from hazard yields more informative evaluation and can improve calibration, with the best joint classification still far from reliable for safety-critical deployment.

What carries the argument

The central object is a two-bit label scheme (hazard × anomaly) that treats danger and contextual irregularity as orthogonal dimensions, formalized as classification codes 00 (Safe), 10 (Anomalous), 01 (Hazardous), 11 (Anomalous-Hazardous). The key mechanism is the prompt structure that explicitly defines "hazard" (OSHA-grounded) and "anomaly" (context violation) and instructs the model to decide each label independently. This orthogonal labeling allows the authors to separate object-level risk cues from contextual deviations, and to measure the over-reaction problem directly.

What would settle it

A direct test would be to run the same four-class evaluation on a set of real photographs of safety-critical scenes (e.g., from construction sites or emergency response footage) with validated ground truth from domain experts, and check whether the measured anomaly/hazard conflation persists or vanishes. If the conflation disappears on real images, the synthetic dataset was the cause.

Watch

Extended reading notes

Core claim

The paper's core claim is that state-of-the-art VLMs conflate anomalousness with hazardousness under binary safety prompting, treating unusual context as if it were physical danger. By forcing models to make two independent binary judgments—one for hazard (potential immediate physical harm) and one for anomaly (violation of expected spatial/contextual relations)—the authors demonstrate that this conflation is measurable and consistent across models and prompts. They find that per-class anomaly F1 scores are consistently lower than hazard F1 scores, and that joint classification (four classes) has a best Hamming loss of 0.23, meaning models still mislabel nearly a quarter of scenes even in th

Load-bearing premise

The study assumes that artificially generated images, created from text descriptions without a specified generation model or validation against real footage, faithfully represent the visual semantics of real hazard and anomaly scenes.

Editorial extensions

If this is right

  • Binary safe/unsafe safety evaluations of VLMs systematically overstate danger in unusual-but-benign scenes, so deployment in high-risk environments could produce false alarms that distract human operators.
  • Explicitly separating anomaly from hazard during prompting improves calibration and yields better per-class performance than a single safety judgment, even when the final decision is collapsed to binary.
  • Dense captions as input do not improve classification over images; they often degrade it, suggesting that current VLMs lose crucial visual context when reasoning from text alone.
  • Zero-shot prompting with the four-class taxonomy significantly outperforms a prior binary prompt on the VERI benchmark (t-test p=0.0022), indicating that the taxonomy itself aids reasoning.
  • No tested model, prompt, or input modality achieves joint four-class accuracy sufficient for safety-critical deployment; the best Hamming loss is still 0.23.

Reading between the lines

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

  • The observed conflation of anomaly and hazard may be a general property of large pretrained models, not limited to these six; if so, any VLM used in situational-awareness tasks should be audited with a four-class rather than binary evaluation.
  • The authors' finding that anomaly classification is consistently harder than hazard classification could be exploited by calibrating decision thresholds differently for the two axes, or by training on more nuanced anomaly examples.
  • The weak agreement among human annotators on borderline cases suggests that even human ground truth for anomaly vs. hazard is not crisp; a probabilistic labeling scheme (e.g., crowd-sourced with consensus measures) might yield a more reliable benchmark and potentially change which models appear best.
  • The paper's implication that forcing a four-class taxonomy improves even binary outcomes hints that richer task decomposition could be a general prompt-engineering strategy for safety judgments, worth testing on other binary safety tasks.
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 introduces a four-class hazard/anomaly taxonomy (Safe, Anomalous, Hazardous, Anomalous-Hazardous) to evaluate whether Vision-Language Models (VLMs) can separate contextual irregularity from physical danger. The authors construct a 610-image dataset, evaluate six VLMs under three prompting strategies (zero-shot, few-shot, chain-of-thought) and two input modalities (images and dense captions), and compare against a binary-safety baseline (VERI/BSTS). The central claim is that VLMs frequently misinterpret anomalousness as hazardousness, and that an explicit four-class scheme exposes failure modes hidden by binary evaluations.

Significance. If substantiated, this is a useful contribution: binary safe/unsafe benchmarks cannot distinguish a model that flags a red costume as blood from one that identifies an actual fire hazard, and the proposed taxonomy is a natural corrective. The paper provides a public dataset and a systematic multi-model, multi-prompt evaluation, which are valuable resources for the VLM safety community. However, the headline claim about anomaly-to-hazard conflation is not directly established by the reported metrics, and several methodological choices (fully synthetic images, unspecified generation pipeline, a baseline comparison that confounds prompt design with taxonomy, and the use of Gemini 3 Flash as both caption generator and evaluated model) weaken the empirical foundation. The work is therefore a promising starting point rather than a definitive demonstration.

major comments (4)
  1. [Abstract / Section IV, Tables I and II] The central claim that 'VLMs frequently misinterpret anomalousness as hazardousness' is not directly supported by the reported results. Table I reports per-class precision/recall/F1 for anomaly and hazard, and Table II reports Hamming loss; neither isolates the error direction. An anomalous image misclassified as 'Both' or 'Hazardous' would lower anomaly recall or joint accuracy, but the aggregate metrics cannot distinguish that specific conflation from other error patterns. The paper needs a confusion matrix (4x4) or conditional error rates such as P(predict Hazard | true Anomaly) and P(predict Anomaly | true Hazard) to demonstrate the claimed over-reliance on contextual irregularity. Without this, the abstract's main conclusion is an interpretation, not an empirical result.
  2. [Section III, Dataset Design; Section V, Limitations] The dataset is described as 'artificially generated based on textual descriptions,' but the text-to-image model, generation prompts, and any validation that the synthetic images faithfully instantiate the intended hazard/anomaly semantics are not specified. The paper acknowledges dataset size and annotation subjectivity but does not address the synthetic-to-real transfer question. If the generated images contain systematic artifacts (e.g., odd rendering that itself reads as 'anomalous'), the measured conflation may not reflect real deployment. Please document the generation pipeline, show example images per class, and provide some evidence (e.g., human agreement on a random subset) that synthetic images match the intended categories.
  3. [Section IV, Baseline Comparison; Figure 4] The comparison to the BSTS baseline confounds two factors: the four-class taxonomy and the prompt wording. The binary mapping ('anomalous'→'safe', 'anomalous-hazardous'→'hazardous') collapses the very distinction the paper studies, and the zero-shot prompt differs from the BSTS prompt in structure and content. The reported significant improvement (p=0.0022) could therefore reflect prompt phrasing alone. A controlled ablation is needed: e.g., evaluate the same four-class prompt but with only binary labels, or keep the prompt identical except for the removal of the anomaly class, to isolate the effect of the taxonomy.
  4. [Section III, Dense Captioning; Section IV, Dense Caption Input] Gemini 3 Flash generated the dense captions and is also one of the evaluated models. Its caption-conditioned scores may be inflated because the model is effectively reasoning over its own text generation, which could encode the same biases. The paper notes the captions were reviewed but does not address this circularity. Please either generate captions with a model that is not in the evaluation set, or explicitly flag and exclude Gemini 3 Flash's caption-conditioned results when interpreting the comparison.
minor comments (5)
  1. [Section III, Dataset Design] The annotation process says researchers 'independently labeled' then 'reviewed collectively to reach a unanimous agreement.' No inter-annotator agreement metric is reported. The limitation is acknowledged in Section V, but a Cohen's kappa or similar statistic would strengthen the dataset's credibility.
  2. [Section III, Formalization] The symbolic framework (P, Q_T_i, Q_I_i, R_i, c_i) is introduced but not used consistently in the rest of the paper. It can be simplified or tied more explicitly to the prompts in the task-design subsections.
  3. [Tables I and II] Table I is dense and lacks a note explaining the subscripts or how 'per-class' treats the 'Both' category. Table II omits a row for Gemini 2.5 Flash with chain-of-thought prompting; the reason ('did not produce valid responses') should be stated in the table caption rather than only in the text.
  4. [Figures 2 and 3] The captions say 'precision vs. recall graph' but do not explain what each point or curve represents (one point per model-prompt combination? a PR curve?). Please clarify the plotting convention.
  5. [References] Reference [9] is dated 2026 and some Roboflow dataset URLs are not standard archival references; please ensure URLs are stable or cite DOIs where possible.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity in the central claim; one minor self-reference exists in the dense-caption sub-task, which is not load-bearing.

  1. other [Section III (Dense Captioning; Model Selection) and Section IV (Table II, Fig. 3)]
    "To generate dense captions, Q T i , for each image of our dataset, we use Google’s Gemini 3 Flash. ... The agent generating the dense captions, Gemini 3 Flash, was given no information about the classification task in order for it to not imbue a bias on other models when they classify anomalies or hazards."

    For Gemini 3 Flash, the dense caption input is itself an output of the same model that is later evaluated on those captions. Its caption-conditioned classification therefore partly measures self-consistency between its own caption generation and its own classification, rather than an independent ability to reason from externally produced text. Table II and Fig. 3 report Gemini 3 Flash’s caption-conditioned scores alongside other models, so this one model’s caption results are partly determined by its own prior output. This affects only the secondary dense-caption modality; the headline anomaly/hazard conflation conclusion is based on image-input results, so the central claim remains independent of this minor self-reference.

full rationale

The paper contains no fitted parameters, no mathematical derivation, and no equation that reduces a predicted quantity to an input. Its central claim is an empirical measurement of VLM outputs on a constructed four-class benchmark, compared against an external baseline ([5]). The self-citations to the authors’ prior work ([12]) are contextual and do not carry the argument. The acknowledged limitations—small per-class sample size, non-expert annotators, and refinement of definitions during annotation—affect generality and label reliability, but they are not circularity. The only mild self-referential element is that Gemini 3 Flash generated the dense captions and is also evaluated on those captions; this does not support the main image-based conclusion. Overall, no significant circularity is present.

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

The paper's measurements rest on several domain assumptions about the realism of synthetic images, the validity of non-expert labels, the neutrality of model-generated captions, and the orthogonality of the hazard/anomaly taxonomy; none of these are independently verified, and the paper's own limitations section concedes the first two.

assumptions (5)
  • ad hoc to paper Text-to-image generation (unspecified model) produces images whose hazard/anomaly content matches real safety-critical scenes.
    Section III Dataset Design states the main dataset comprises artificially generated images based on textual descriptions; no generation model or validation against real images is provided, so all VLM measurements inherit this assumption.
  • domain assumption OSHA-based definition of hazard and the paper's definition of anomaly are orthogonal and exhaustively cover the four classes.
    Section III Dataset Design defines two-bit labels; orthogonality and exhaustiveness are asserted, not empirically validated.
  • domain assumption Unanimous agreement among a few non-expert annotators after discussion yields reliable ground-truth labels.
    Section III Dataset Design describes the annotation process; Section V Limitations concedes subjectivity and lack of consensus measures.
  • domain assumption Dense captions generated by Gemini 3 Flash and reviewed by the authors are neutral and sufficiently accurate for classification.
    Section III Dense Captioning; captions were authored by a model that is also evaluated, and review was done by the labelers themselves.
  • domain assumption Excluding model outputs that violate the strict format does not bias the measured performance.
    Section IV and figure captions omit conditions with more than 10 invalid responses; the paper does not test whether invalid-response frequency correlates with class or model capability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hazard or Anomaly? Evaluating VLMs for Understanding Dangers and Discrepancies." pith.science (2026). https://pith.science/paper/DFSR5VHP

@misc{pith2026260718325,
  author       = {Pith},
  title        = {Pith review of: Hazard or Anomaly? Evaluating VLMs for Understanding Dangers and Discrepancies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DFSR5VHP}},
  note         = {Machine review of arXiv:2607.18325}
}
read the original abstract

Modern safety-critical systems increasingly rely on human-robot interaction to reduce disaster risk and support decision-making during emergencies. Vision-Language Models (VLMs) are promising for these settings because they can interpret complex scenes and communicate safety-relevant information, but they still require careful evaluation to ensure reliable safety reasoning. In particular, current evaluations often frame danger recognition as a binary decision (Safe/Unsafe), making it unclear whether a model is identifying true physical hazards or merely reacting to unusual scene elements. We address this limitation by introducing an explicit distinction between hazard and anomaly, and by separately recognizing hazardous and anomalous states. We evaluate several state-of-the-art VLMs across two datasets and multiple prompting strategies to test whether this distinction changes model behavior. Our results show that VLMs frequently misinterpret anomalousness as hazardousness, revealing an over-reliance on contextual irregularity as a proxy for danger. We further show that explicitly separating anomaly from hazard provides a more informative evaluation of VLM safety reasoning and exposes failure modes that binary safety judgments can obscure. Our public dataset is available on Roboflow https://app.roboflow.com/vlm-in-context-anomaly-and-hazard-detection/camera-ready-roman-ds.

Figures

Figures reproduced from arXiv: 2607.18325 by the authors.

Figure 1
Figure 1. Images extracted from our public dataset representing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The precision vs. recall graph when using images as [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Shows how much models underpredicted or overpredicted images classified as “safe" or “dangerous” extracted from [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 4 linked inside Pith

  1. [1]

    COCO-OOC dataset,

    M. Acharya, A. Roy, and S. Jha, “COCO-OOC dataset,” Apr. 2024

    2024

  2. [2]

    Why context matters in VQA and Reasoning: Semantic interventions for VLM input modalities,

    K. Amara, L. Klein, C. Lüth, P. Jäger, H. Strobelt, and M. El-Assady, “Why context matters in VQA and Reasoning: Semantic interventions for VLM input modalities,” Oct. 2024

    2024

  3. [3]

    Chandhok

    S. Chandhok. SceneGPT: A Language Model for 3D Scene Understanding. [Online]. Available: http://arxiv.org/abs/2408.06926

    arXiv

  4. [4]

    Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas

    S. Chen, T. Zhu, R. Zhou, J. Zhang, S. Gao, J. C. Niebles, M. Geva, J. He, J. Wu, and M. Li, “Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas.”

  5. [5]

    Better safe than sorry? overreaction problem of vision language models in visual emergency recognition,

    D. Choi, S. Lee, and Y . Song, “Better safe than sorry? overreaction problem of vision language models in visual emergency recognition,” arXiv preprint arXiv:2505.15367, 2025

    arXiv 2025

  6. [6]

    Real-time robotics situation awareness for accident prevention in industry,

    J. M. Deniz, A. S. Kelboucas, and R. B. Grando, “Real-time robotics situation awareness for accident prevention in industry,” 2024

    2024

  7. [7]

    Multilabel Prediction with Probability Sets: The Ham- ming Loss Case,

    S. Destercke, “Multilabel Prediction with Probability Sets: The Ham- ming Loss Case,” inInformation Processing and Management of Un- certainty in Knowledge-Based Systems. Cham: Springer International Publishing, 2014, vol. 443, pp. 496–505

    2014

  8. [8]

    Smart operators: How industry 4.0 is affecting the worker’s performance in manufacturing contexts,

    V . Di Pasquale, V . De Simone, S. Miranda, and S. Riemma, “Smart operators: How industry 4.0 is affecting the worker’s performance in manufacturing contexts,”Procedia Computer Science, vol. 180, pp. 958–967, 2021

    2021

Show all 42 references
  1. [9]

    Roboflow,

    B. Dwyer, J. Nelson, T. Hansenet al., “Roboflow,” 2026, computer Vision Management Platform. [Online]. Available: https: //roboflow.com

    2026

  2. [10]

    Semantic anomaly detection with large language models,

    A. Elhafsi, R. Sinha, C. Agia, E. Schmerling, I. A. D. Nesnas, and M. Pavone, “Semantic anomaly detection with large language models,” Autonomous Robots, vol. 47, no. 8, pp. 1035–1055, Dec. 2023

    2023

  3. [11]

    Situation awareness: State of the art,

    M. R. Endsley and E. S. Connors, “Situation awareness: State of the art,” in2008 IEEE Power and Energy Society General Meeting - Conversion and Delivery of Electrical Energy in the 21st Century. Pittsburgh, PA, USA: IEEE, July 2008, pp. 1–4

    2008

  4. [12]

    LLM-supported safety annotation in high-risk environments,

    M. Eskandari, M. K. V . Indukuri, S. M. Lukin, and C. Matuszek, “LLM-supported safety annotation in high-risk environments,” in Proceedings of the HRI 2025 Workshop on Vision-Based Assistance for Manipulation in Human-Robot Interaction (VAM-HRI), 2025. [Online]. Available: htt...

    2025

  5. [13]

    Understanding human behaviour in industrial human– robot interaction by means of virtual reality,

    P. Fratczak, Y . M. Goh, P. Kinnell, A. Soltoggio, and L. Justham, “Understanding human behaviour in industrial human– robot interaction by means of virtual reality,” inProceedings of the Halfway to the Future Symposium 2019. Nottingham, United Kingdom: ACM, 2019, pp. 1–7. [On...

    2019

  6. [14]

    interior computer vision dataset,

    grad, “interior computer vision dataset,” 2024. [Online]. Available: https://universe.roboflow.com/grad-ltvg6/interior-umkr0

    2024

  7. [15]

    When hci meets hri: The intersection and distinction,

    W. Huang, “When hci meets hri: The intersection and distinction,” Virginia Polytechnic Institute and State University, 2015

    2015

  8. [16]

    Safety hazard identification computer vision model,

    A. S. H. Identifier, “Safety hazard identification computer vision model,” 2025. [Online]. Available: https://universe.roboflow.com/ ai-safety-hazard-identifier/safety-hazard-identification

    2025

  9. [17]

    The role of data and information quality during disaster response decision- making,

    V . Jayawardene, T. Huggins, R. Prasanna, and B. Fakhruddin, “The role of data and information quality during disaster response decision- making,”Progress in Disaster Science, vol. 12, p. 100202, 08 2021

    2021

  10. [18]

    Jentsch,Human-Robot Interactions in Future Military Operations

    F. Jentsch,Human-Robot Interactions in Future Military Operations. CRC Press, 2016

    2016

  11. [19]

    VRU-accident: A vision-language benchmark for video question answering and dense captioning for accident scene understanding,

    Y . Kim, A. S. Abdelrahman, and M. Abdel-Aty, “VRU-accident: A vision-language benchmark for video question answering and dense captioning for accident scene understanding,” 2025

    2025

  12. [20]

    firefighter computer vision dataset,

    J. Ko, “firefighter computer vision dataset,” 2022. [Online]. Available: https://universe.roboflow.com/juyeon-ko/firefighter

    2022

  13. [21]

    aerial computer vision model,

    krauseswelt, “aerial computer vision model,” 2024. [Online]. Available: https://universe.roboflow.com/krauseswelt/aerial-qjpyp

    2024

  14. [22]

    Large Language Models and Multimodal Retrieval for Visual Word Sense Disambigua- tion,

    A. Kritharoula, M. Lymperaiou, and G. Stamou, “Large Language Models and Multimodal Retrieval for Visual Word Sense Disambigua- tion,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 13 053–13 077

    2023

  15. [24]

    Information overload, stress, and emergency managerial thinking,

    S. Misra, P. Roberts, and M. Rhodes, “Information overload, stress, and emergency managerial thinking,”International Journal of Disaster Risk Reduction, vol. 46, p. 101615, 2020. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S2212420920312644

    2020

  16. [25]

    Natural disaster damage computer vision model,

    model v2, “Natural disaster damage computer vision model,”

  17. [26]

    Chemical spill computer vision model,

    Nhyan, “Chemical spill computer vision model,” 2025. [Online]. Available: https://universe.roboflow.com/nhyan/chemical-spill-b5dmp

    2025

  18. [27]

    Hazard assessment and job safety analysis,

    Occupational Safety and Health Administration (OSHA), “Hazard assessment and job safety analysis,” 2018, u.S. Department of Labor, Occupational Safety and Health Administration

    2018

  19. [28]

    Virtual reality based space operations – a study of esa’s potential for vr based training and simulation,

    M. Olbrich, H. Graf, J. Keil, R. Gad, S. Bamfaste, and F. Nicolini, “Virtual reality based space operations – a study of esa’s potential for vr based training and simulation,” inVirtual, Augmented and Mixed Reality: Interaction, Navigation, Visualization, Embodiment, and Simul...

    2018

  20. [29]

    Vision language models are blind,

    P. Rahmanzadehgervi, L. Bolton, M. R. Taesiri, and A. T. Nguyen, “Vision language models are blind,” inProceedings of the Asian Conference on Computer Vision (ACCV), December 2024, pp. 18–34

    2024

  21. [30]

    MSTS: A Multimodal Safety Test Suite for Vision-Language Models,

    P. Röttger, G. Attanasio, F. Friedrich, J. Goldzycher, A. Parrish, R. Bhardwaj, C. D. Bonaventura, R. Eng, G. E. K. Geagea, S. Goswami, J. Han, D. Hovy, S. Jeong, P. Jereti ˇc, F. M. Plaza-del- Arco, D. Rooein, P. Schramowski, A. Shaitarova, X. Shen, R. Willats, A. Zugarini, a...

    2025

  22. [31]

    Classification ppe computer vision model,

    S. D. B. SHAFIZAL, “Classification ppe computer vision model,” 2025. [Online]. Available: https://universe.roboflow.com/ shafia-diana-binti-shafizal/classification-ppe

    2025

  23. [32]

    Assessing GPT’s Potential for Word Sense Disambiguation: A Quantitative Evaluation on Prompt Engineering Techniques,

    D. Sumanathilaka, N. Micallef, and J. Hough, “Assessing GPT’s Potential for Word Sense Disambiguation: A Quantitative Evaluation on Prompt Engineering Techniques,” in2024 IEEE 15th Control and System Graduate Research Colloquium (ICSGRC), Aug. 2024, pp. 204–209

    2024

  24. [33]

    Winoground: Probing vision and language models for visio-linguistic compositionality,

    T. Thrush, R. Jiang, M. Bartolo, A. Singh, A. Williams, D. Kiela, and C. Ross, “Winoground: Probing vision and language models for visio-linguistic compositionality,” 2022, version Number: 2. [Online]. Available: https://arxiv.org/abs/2204.03162

    2022 arXiv

  25. [34]

    ConTextual: Evaluating Context-Sensitive Text-Rich Visual Reasoning in Large Multimodal Models,

    R. Wadhawan, H. Bansal, K.-W. Chang, and N. Peng, “ConTextual: Evaluating Context-Sensitive Text-Rich Visual Reasoning in Large Multimodal Models,” Jul. 2024

    2024

  26. [35]

    Caption This, Reason That: VLMs Caught in the Middle,

    Z. Weng, L. Gomez, T. W. Webb, and P. Bashivan, “Caption This, Reason That: VLMs Caught in the Middle,” May 2025

    2025

  27. [36]

    HazardVLM: A Video Language Model for Real-Time Hazard Description in Auto- mated Driving Systems,

    D. Xiao, M. Dianati, P. Jennings, and R. Woodman, “HazardVLM: A Video Language Model for Real-Time Hazard Description in Auto- mated Driving Systems,”IEEE Transactions on Intelligent V ehicles, vol. 10, no. 5, pp. 3331–3343, May 2025

    2025

  28. [37]

    VLM: Task-agnostic Video- Language Model Pre-training for Video Understanding,

    H. Xu, G. Ghosh, P.-Y . Huang, P. Arora, M. Aminzadeh, C. Feicht- enhofer, F. Metze, and L. Zettlemoyer, “VLM: Task-agnostic Video- Language Model Pre-training for Video Understanding,” Sep. 2021

    2021

  29. [38]

    Leveraging large language models for word sense disambiguation,

    J. H. Yae, N. C. Skelly, N. C. Ranly, and P. M. LaCasse, “Leveraging large language models for word sense disambiguation,”Neural Com- puting and Applications, vol. 37, no. 6, pp. 4093–4110, Feb. 2025

    2025

  30. [39]

    Common inpainted objects in-n-out of context,

    T. Yang, T. Jordan, N. Liu, and J. Sun, “Common inpainted objects in-n-out of context,”arXiv:2506.00721, 2025. [Online]. Available: https://arxiv.org/abs/2506.00721

    2025 arXiv

  31. [40]

    Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi,

    X. Yue, Y . Ni, K. Zhang, T. Zheng, R. Liu, G. Zhang, S. Stevens, D. Jiang, W. Ren, Y . Sunet al., “Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2024

  32. [41]

    Do llms understand visual anomalies? Uncovering llm’s capabilities in zero-shot anomaly detection,

    J. Zhu, S. Cai, F. Deng, B. C. Ooi, and J. Wu, “Do llms understand visual anomalies? Uncovering llm’s capabilities in zero-shot anomaly detection,” inProceedings of the 32nd ACM International Conference on Multimedia, ser. Mm ’24. New York, NY , USA: Association for Computing ...

    2024

  33. [2023]

    Available: https://universe.roboflow.com/model-v2/ natural-disaster-damage-np4mh

    [Online]. Available: https://universe.roboflow.com/model-v2/ natural-disaster-damage-np4mh

  34. [2025]

    Available: https://arxiv.org/abs/2506.05405

    [Online]. Available: https://arxiv.org/abs/2506.05405

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.