Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Understanding and Mitigating Toxicity in Image-Text Pretraining Datasets: A Case Study on LLaVA

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A three-detector pipeline removes 7,531 toxic image-text pairs from LLaVA's pretraining set.

desk verdict Useful dataset artifact, but the 7,531 'toxic' claim rests on unvalidated detectors and no downstream test; worth reviewing but needs validation. read the letter →

arxiv 2505.06356 v1 pith:S66X43R6 submitted 2025-05-09 cs.CV

Karthik Reddy Kanjula , Surya Guthikonda , Nahid Alam , Shayekh Bin Islam This is my paper · ORCID
classification cs.CV
keywords toxicitydetectionmultimodalsafetyLLaVAimage-textpretrainingdatasetfilteringLlavaGuardToxic-BERTresponsibleAI
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

Pretraining datasets assembled from the web carry harmful images and captions, and models trained on them can absorb that content. This paper claims that LLaVA's widely used 558,000-pair pretraining set contains measurable toxicity, and that a multimodal filtering pipeline can find and remove it. Combining an image-safety model (LlavaGuard) with a second-pass reranker (Command R+) and a text-toxicity model (Toxic-BERT), the paper flags 7,531 unique toxic image-text pairs, roughly 1.35 percent of the dataset, and removes them to produce an openly released, toxicity-mitigated version of the pretraining data. If the detection is accurate, the result offers a practical recipe for cleaning web-scale multimodal data before training.

What carries the argument

The load-bearing mechanism is a two-modality, three-stage filtering pipeline. LlavaGuard 7B is a vision-language safety classifier that rates each image against nine policy categories (hate/harassment, violence, sexual content, nudity, criminal planning, weapons/substance abuse, self-harm, animal cruelty, and disasters/emergencies); Command R+ is a large language model that re-examines each LlavaGuard-flagged image with an optimized prompt to cut false positives; Toxic-BERT is a text classifier fine-tuned to score captions for toxic language and flag those above 0.8 confidence. The pipeline then unions the surviving image flags with the caption flags and removes every unique flagged pair.

What would settle it

Take a random sample of, say, 200 of the 7,531 removed pairs and 200 retained pairs, mask the source, and have independent annotators label each pair as toxic or not; if the removed pairs are not clearly more toxic than the retained pairs, or if a large share of removed pairs are benign, then the 7,531 count and the 'mitigated' label are not established.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that toxicity in LLaVA pretraining is detectable and removable with off-the-shelf components: LlavaGuard flags 7,600 images under its nine-category safety taxonomy, Command R+ narrows that to 7,111 after a prompt-tuned re-evaluation, and Toxic-BERT flags 892 captions at over 80 percent confidence; the union is 7,531 unique pairs. The distribution is dominated by sexual content (33.1%), weapons or substance abuse (22.0%), nudity (19.5%), and violence, harm, or cruelty (16.3%). Removing these pairs yields a toxicity-mitigated LLaVA pretraining dataset, which the paper releases open source.

Load-bearing premise

The whole removal count depends on treating the three detectors' labels as accurate measures of toxicity, since no human verification or independent benchmark is used to confirm them.

Editorial extensions

If this is right

  • A public toxicity-mitigated LLaVA pretraining set now exists, with 7,531 flagged pairs removed, for anyone to use in pretraining.
  • The same pipeline can be applied to other image-text pretraining collections without retraining detectors, since all three components are off the shelf.
  • The category breakdown gives dataset curators specific targets: sexual content and weapons or substance abuse account for more than half of the flagged images.
  • Because only about 1.35 percent of the 558,000 pairs are removed, the filtered set preserves almost all of the original data's volume and the paper's stated richness and diversity.

Reading between the lines

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

  • Editorial inference: the 7,531 count reflects these specific detectors' criteria, not a verified ground truth, so a human audit would likely shift the number in either direction.
  • Editorial inference: the arithmetic overlap between the image-only flags (7,111) and caption-only flags (892) is only about 472 pairs, suggesting image toxicity and text toxicity are largely disjoint and that both modalities need separate filters.
  • Editorial inference: a natural testable extension is to pretrain a LLaVA-style model on the mitigated set and compare safety-benchmark behavior plus standard VQA or captioning accuracy against a model trained on the original set; the paper defers that evaluation to future work.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper analyzes the LLaVA image-text pretraining dataset (558,000 pairs) for toxic content using a multimodal pipeline: LlavaGuard 7B flags potentially unsafe images (7,600), Command R+ with a prompt tuned via Cohere's prompt tuner refines this set to 7,111, and Toxic-BERT flags 892 captions above a hand-set 0.8 confidence threshold. The authors take the union, 7,531 unique image-text pairs, remove them, and release a 'toxicity-mitigated' version of the dataset. The paper also reports the distribution of LlavaGuard toxicity categories and Toxic-BERT toxicity types, and outlines future plans for safety evaluation.

Significance. If the claims are validated, the paper would provide a useful open resource: a filtered version of the widely used LLaVA pretraining dataset, plus an analysis of where toxicity appears across visual and textual modalities. The pipeline is clearly described and the release is a concrete community contribution. However, the significance is conditional on the accuracy of the detectors; the paper provides no evidence that the 7,531 flagged pairs are actually toxic by any external standard. The manuscript itself acknowledges in Section 5 that verifying the accuracy of the dataset is future work, which is a serious gap for the central claim.

major comments (3)
  1. [Section 3.2, Figure 3] The central quantitative claim—that 7,531 image-text pairs are 'toxic' and their removal produces a 'toxicity-mitigated' dataset—rests entirely on the outputs of three models, with no human verification, no external toxicity benchmark, and no precision/recall or false-positive analysis for any of the detectors on this specific data distribution. LlavaGuard was trained for general VLM safety and Toxic-BERT on Jigsaw forum comments, so their behavior on web-scraped LLaVA pretraining pairs is unknown. Because the paper's own Section 5 states that verifying the accuracy of the dataset is future work, the 7,531 count is a procedural assertion ('we ran these models and removed their union') rather than a demonstrated safety property.
  2. [Section 3.2, Command R+ refinement] The refinement step in which Command R+ identifies 'truly unsafe' image IDs from LlavaGuard's 7,600 flags uses an optimized prompt built with Cohere's prompt tuner, but the prompt itself is never disclosed and no independent ground truth is used to validate Command R+'s binary judgments. The reduction from 7,600 to 7,111 images could reflect either removal of LlavaGuard false positives or introduction of Command R+ false negatives; without the prompt and a validation set, the step is neither reproducible nor scientifically assessable. This is load-bearing because the final count of 7,111 depends entirely on this unvalidated model judgment.
  3. [Sections 4 and 6] The paper claims that removing 7,531 pairs yields a 'toxicity-mitigated version for LLaVA pretraining,' but provides no downstream evaluation showing that filtering 1.35% of the dataset preserves model capabilities or actually reduces toxic generations. No comparisons on standard benchmarks (e.g., MM-SafetyBench, VHELM, ELITE, which the paper cites as future work) are reported. Without such evaluation, the practical benefit of the released dataset is unquantified, and the removal could in principle degrade the diversity or quality of the pretraining data without any measurable safety gain.
minor comments (5)
  1. [Throughout] There are numerous typographical artifacts: 'LLaV A' appears with an extra space in the title and abstract, 'T oxicity' and 'W e' appear with stray spaces in Figures 1 and 2 and the abstract, and the author footnote 'Work does not belong to position referred in 2' is unclear.
  2. [Section 3.2 and Figure 2] The caption of Figure 2 reads 'n > 0.8: 892' etc., but 'n' is not defined; it presumably denotes the Toxic-BERT confidence score. Please clarify the notation and also state the total number of captions processed.
  3. [Section 3.2] The paper reports the union of 7,111 image-based flags and 892 text-based flags as 7,531 unique images, but it does not report the overlap between the two sets. Without the contingency table, readers cannot assess how many pairs were flagged by both modalities or whether the multimodal combination actually contributes beyond either detector alone.
  4. [References] Some references are incomplete or inconsistently formatted: the Cohere Command R reference lacks a year and a stable URL, and the Detoxify entry is a GitHub link rather than a citable paper. Also, reference [20] in Section 5 is cited as 'Howard et al.' but the reference is by Ratzlaff et al.
  5. [Section 3.1] The statement that models 'provided a reliable foundation for systematically identifying, understanding, and mitigating toxicity' is not supported by any quantitative reliability measures in this paper, and should be softened or backed by validation results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the 7,531 count is a procedural filter output, not a derivation that reduces to its inputs.

full rationale

This paper does not present a mathematical derivation; it describes an empirical filtering pipeline and reports the resulting count. The central number 7,531 is defined procedurally as the union of three detector outputs: LlavaGuard plus Command R+ yield 7,111 images and Toxic-BERT yields 892 captions. The paper does not define 'toxic' as 'flagged by these models'; Section 2 gives a conceptual definition ('harmful or offensive content'), and the detectors are presented as tools for operationalizing that concept. No equation-level reduction is present, and no fitted parameter is renamed as a prediction: the 0.8 Toxic-BERT threshold and the optimized Command R+ prompt are filter settings, not fitted predictions of an external quantity. The paper's own Future Work section acknowledges that the accuracy of the toxicity-mitigated dataset is not yet verified ('It would be interesting to verify the accuracy of the toxicity-mitigated dataset by applying a user evaluation process or through other toxicity-mitigation pipelines'). That missing validation is a correctness or evidentiary gap, not circularity. Similarly, the use of Cohere's Command R+ and prompt tuner by authors affiliated with Cohere for AI is a potential conflict-of-interest or self-referential tool choice, but it is not a citation chain invoked to prove a result; it is a model invocation. Because no claim reduces by construction to its own inputs, and no load-bearing argument depends on an unverified self-citation, the appropriate finding is no significant circularity.

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

The central numbers rest entirely on three detectors and an undisclosed tuned prompt. No human-verified labels or external benchmark are provided, so the counts are definitional to the pipeline rather than independently established measurements.

free parameters (2)
  • Toxic-BERT confidence threshold = 0.8
    Captions with a toxicity score above 0.8 are flagged as toxic. The threshold is chosen by hand and directly determines the 892-caption count in Figure 2.
  • Command R+ refinement prompt = not disclosed
    A prompt optimized with the Cohere prompt tuner is used to turn LlavaGuard flags into final unsafe labels. The exact prompt and tuning data are not given, yet this drives the 7,111-image count.
assumptions (4)
  • domain assumption LlavaGuard's safety taxonomy and labels correctly identify visual toxicity.
    The paper trusts LlavaGuard's nine-category safety ratings as the starting point for all image filtering, with no validation against human labels.
  • domain assumption Toxic-BERT's contextual toxicity scores correctly identify toxic captions.
    Toxic-BERT is applied directly to captions and the top-scoring 892 are accepted as toxic without manual review.
  • ad hoc to paper Command R+ with the optimized prompt correctly separates true positives from LlavaGuard false positives.
    This judge step is introduced specifically for this pipeline; its accuracy is assumed and never measured.
  • domain assumption Removing the flagged pairs reduces harmful content without materially harming the pretraining dataset.
    The paper calls the result 'toxicity-mitigated' but reports no downstream model evaluation on safety or capability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding and Mitigating Toxicity in Image-Text Pretraining Datasets: A Case Study on LLaVA." pith.science (2026). https://pith.science/paper/S66X43R6

@misc{pith2026250506356,
  author       = {Pith},
  title        = {Pith review of: Understanding and Mitigating Toxicity in Image-Text Pretraining Datasets: A Case Study on LLaVA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S66X43R6}},
  note         = {Machine review of arXiv:2505.06356}
}
read the original abstract

Pretraining datasets are foundational to the development of multimodal models, yet they often have inherent biases and toxic content from the web-scale corpora they are sourced from. In this paper, we investigate the prevalence of toxicity in LLaVA image-text pretraining dataset, examining how harmful content manifests in different modalities. We present a comprehensive analysis of common toxicity categories and propose targeted mitigation strategies, resulting in the creation of a refined toxicity-mitigated dataset. This dataset removes 7,531 of toxic image-text pairs in the LLaVA pre-training dataset. We offer guidelines for implementing robust toxicity detection pipelines. Our findings underscore the need to actively identify and filter toxic content - such as hate speech, explicit imagery, and targeted harassment - to build more responsible and equitable multimodal systems. The toxicity-mitigated dataset is open source and is available for further research.

Figures

Figures reproduced from arXiv: 2505.06356 by the authors.

Figure 1
Figure 1. Image Toxicity Analysis on LLaVA Pre-train Datase [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Image Caption Toxicity Analysis on LLaVA Pre-trai [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 20 canonical work pages

  1. [1]

    Flamingo: a Visual Language Model for Few-Shot Learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Anto ine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a Visual Language Model for Few-Shot Learning. Advances in Neural Information Processing Systems, 35:23716–23736, 2022. 1

    work page 2022

  2. [2]

    Y our vision-language model itself is a strong filter: Toward s high-quality instruction tuning with data selection, 2024

    Ruibo Chen, Yihan Wu, Lichang Chen, Guodong Liu, Qi He, Tianyi Xiong, Chenxi Liu, Junfeng Guo, and Heng Huang. Y our vision-language model itself is a strong filter: Toward s high-quality instruction tuning with data selection, 2024 . 2

    work page 2024

  3. [3]

    Comm: A coherent inter- leaved image-text dataset for multimodal understanding an d generation, 2024

    Wei Chen, Lin Li, Y ongqi Yang, Bin Wen, Fan Yang, Tingt- ing Gao, Y u Wu, and Long Chen. Comm: A coherent inter- leaved image-text dataset for multimodal understanding an d generation, 2024. 1

    work page 2024

  4. [4]

    PaLI: A Jointly- Scaled Multilingual Language-Image Model

    Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni , Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. PaLI: A Jointly- Scaled Multilingual Language-Image Model. arXiv preprint arXiv:2209.06794, 2022. 1

    arXiv 2022

  5. [5]

    PaLI-X: On Scaling up a Multilingual Vision and Language Model

    Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa , Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Se- bastian Goodman, Xiao Wang, Yi Tay, et al. PaLI-X: On Scaling up a Multilingual Vision and Language Model. arXiv preprint arXiv:2305.18565, 2023. 1

    arXiv 2023

  6. [6]

    Command R

    Cohere. Command R. https://cohere.com/command, 2024. 1, 3

    work page 2024

  7. [7]

    Molmo and PixMo: Open Weights and Open Data for State-of-the- Art Multimodal Models

    Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tri- pathi, Y ue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and PixMo: Open Weights and Open Data for State-of-the- Art Multimodal Models. arXiv preprint arXiv:2409.17146 , 2024. 1

    arXiv 2024

  8. [8]

    Detoxify

    Laura Hanu and Unitary team. Detoxify. Github. https://github.com/unitaryai/detoxify, 2020. 1, 2

    work page 2020

Show all 30 references
  1. [9]

    LLavaGuard: VLM-based Safeguards for Vision Dataset Curation and Safety Assess- ment

    Lukas Helff, Felix Friedrich, Manuel Brack, Kristian Ke r- sting, and Patrick Schramowski. LLavaGuard: VLM-based Safeguards for Vision Dataset Curation and Safety Assess- ment. arXiv preprint arXiv:2406.05113, 2024. 1, 2

    2024 arXiv

  2. [10]

    Jailbreakzoo: Survey, landscapes, and horizons in jailbreaking large language an d vision-language models, 2024

    Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, and Haohan Wang. Jailbreakzoo: Survey, landscapes, and horizons in jailbreaking large language an d vision-language models, 2024. 1

    2024

  3. [11]

    Vhelm: A holistic evaluation of vision language mod- els, 2024

    Tony Lee, Haoqin Tu, Chi Heem Wong, Wenhao Zheng, Yiyang Zhou, Yifan Mai, Josselin Somerville Roberts, Michihiro Yasunaga, Huaxiu Yao, Cihang Xie, and Percy Liang. Vhelm: A holistic evaluation of vision language mod- els, 2024. 2, 4

    2024

  4. [12]

    Elite: Enhanced language-image toxicity evaluation for safety, 2025

    Wonjun Lee, Doehyeon Lee, Eugene Choi, Sangyoon Y u, Ashkan Y ousefpour, Haon Park, Bumsub Ham, and Suhyun Kim. Elite: Enhanced language-image toxicity evaluation for safety, 2025. 2, 4

    2025

  5. [13]

    Improved Baselines with Visual Instruction Tuning, 2023

    Haotian Liu, Chunyuan Li, Y uheng Li, and Y ong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2023. 1, 2

    2023

  6. [14]

    Visual Instruction Tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Y ong Jae Lee. Visual Instruction Tuning, 2023. 1, 2

    2023

  7. [15]

    Mm-safetybench: A benchmark for safety eval- uation of multimodal large language models, 2024

    Xin Liu, Yichen Zhu, Jindong Gu, Y unshi Lan, Chao Yang, and Y u Qiao. Mm-safetybench: A benchmark for safety eval- uation of multimodal large language models, 2024. 2, 4

    2024

  8. [16]

    Towards interpreting visual infor - mation processing in vision-language models, 2024

    Clement Neo, Luke Ong, Philip Torr, Mor Geva, David Krueger, and Fazl Barez. Towards interpreting visual infor - mation processing in vision-language models, 2024. 1

    2024

  9. [17]

    Kosmos-G: Generating Images in Context with Multimodal Large Language Models

    Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-G: Generating Images in Context with Multimodal Large Language Models. ArXiv, abs/2310.02992, 2023. 1

    2023 arXiv

  10. [18]

    Kosmos-2: Grounding Multimodal Large Language Models to the World

    Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding Multimodal Large Language Models to the World. arXiv preprint arXiv:2306.14824, 2023. 1

    2023 arXiv

  11. [19]

    Learn- ing Transferable Visual Models From Natural Language Su- pervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing Transferable Visual Models From Natural Language Su- pervision. In International conference on machine learning...

    2021

  12. [20]

    Training-free mitigation of language reasoning degradation after multimodal instruction tuning, 2024

    Neale Ratzlaff, Man Luo, Xin Su, V asudev Lal, and Philli p Howard. Training-free mitigation of language reasoning degradation after multimodal instruction tuning, 2024. 4

    2024

  13. [21]

    Laion-5b: An open large-scale dataset for training next generation image-text models, 2022

    Christoph Schuhmann, Romain Beaumont, Richard V encu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. La...

    2022

  14. [22]

    From pixels to prose: A large dataset of dense image cap- tions, 2024

    V asu Singla, Kaiyu Y ue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image cap- tions, 2024. 1, 2

    2024

  15. [23]

    RoFormer: Enhanced Transformer with Rotary Position Embedding, 2021

    Jianlin Su, Y u Lu, Shengfeng Pan, Bo Wen, and Y unfeng Liu. RoFormer: Enhanced Transformer with Rotary Position Embedding, 2021. 1

    2021

  16. [24]

    Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan , Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 1

    2024 arXiv

  17. [25]

    Florence-2: Advancing a unified representation for a variet y of vision tasks

    Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Y umao Lu, Michael Zeng, Ce Liu, and Lu Y uan. Florence-2: Advancing a unified representation for a variet y of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4...

    2024

  18. [26]

    Pangea: A Fully Open Multilin- gual Multimodal LLM for 39 Languages

    Xiang Y ue, Y ueqi Song, Akari Asai, Seungone Kim, Jean de Dieu Nyandwi, Simran Khanuja, Anjali Kan- tharuban, Lintang Sutawika, Sathyanarayanan Ramamoor- thy, and Graham Neubig. Pangea: A Fully Open Multilin- gual Multimodal LLM for 39 Languages. arXiv preprint arXiv:2410.161...

    arXiv

  19. [27]

    Sigmoid Loss for Language Image Pre- Training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, an d Lucas Beyer. Sigmoid Loss for Language Image Pre- Training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975–11986, 2023. 1

    2023

  20. [28]

    Spa-vl: A comprehensive safety preference alignment dataset for vi - sion language model, 2025

    Y ongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Y u Qiao, Xuan- jing Huang, Feng Zhao, Tao Gui, and Jing Shao. Spa-vl: A comprehensive safety preference alignment dataset for vi - sion language model, 2025. 2, 4

    2025

  21. [29]

    Zero-shot defense against toxic images via inherent multimodal alignment in lvlms, 2025

    Wei Zhao, Zhe Li, Yige Li, and Jun Sun. Zero-shot defense against toxic images via inherent multimodal alignment in lvlms, 2025. 2, 4

    2025

  22. [30]

    Un- derstanding and rectifying safety perception distortion i n vlms, 2025

    Xiaohan Zou, Jian Kang, George Kesidis, and Lu Lin. Un- derstanding and rectifying safety perception distortion i n vlms, 2025. 2

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.