MORALISE: A Structured Benchmark for Moral Alignment in Visual Language Models

Xiao Lin, Zhining Liu, Ze Yang, Gaotang Li, Ruizhong Qiu, Shuke Wang, Hui Liu, Haotian Li, Yuchen Yan, Sumit Keswani, Vishwa Pardeshi, Huijun Zhao, Wei Fan, Hanghang Tong
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:74723-74744, 2026.

Abstract

Recently, vision-language models have demonstrated increasing influence in morally sensitive domains such as autonomous driving and medical analysis, owing to their powerful multimodal reasoning capabilities. As these models are deployed in high-stakes real-world applications, it is of paramount importance to ensure that their outputs align with human moral values and remain within moral boundaries. However, existing work on moral alignment either focuses solely on textual modalities or relies heavily on AI-generated images, leading to distributional biases and reduced realism. To overcome these limitations, we introduce MORALISE, a comprehensive benchmark for evaluating the moral alignment of vision-language models (VLMs) using diverse, expert-verified real-world data. We begin by proposing a comprehensive taxonomy of 13 moral topics grounded in Turiel’s Domain Theory, spanning the personal, interpersonal, and societal moral domains encountered in everyday life. Built on this framework, we manually curate 2,481 high-quality image-text pairs, each annotated with two fine-grained labels: (1) topic annotation, identifying the violated moral topic(s), and (2) modality annotation, indicating whether the violation arises from the image or the text. For evaluation, we encompass two tasks, moral judgment and moral norm attribution, to assess models’ awareness of moral violations and their reasoning ability on morally salient content. Extensive experiments on 19 popular open- and closed-source VLMs show that MORALISE poses a significant challenge, revealing persistent moral limitations in current state-of-the-art models.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-lin26aa, title = {{MORALISE}: A Structured Benchmark for Moral Alignment in Visual Language Models}, author = {Lin, Xiao and Liu, Zhining and Yang, Ze and Li, Gaotang and Qiu, Ruizhong and Wang, Shuke and Liu, Hui and Li, Haotian and Yan, Yuchen and Keswani, Sumit and Pardeshi, Vishwa and Zhao, Huijun and Fan, Wei and Tong, Hanghang}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {74723--74744}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/lin26aa/lin26aa.pdf}, url = {https://proceedings.mlr.press/v306/lin26aa.html}, abstract = {Recently, vision-language models have demonstrated increasing influence in morally sensitive domains such as autonomous driving and medical analysis, owing to their powerful multimodal reasoning capabilities. As these models are deployed in high-stakes real-world applications, it is of paramount importance to ensure that their outputs align with human moral values and remain within moral boundaries. However, existing work on moral alignment either focuses solely on textual modalities or relies heavily on AI-generated images, leading to distributional biases and reduced realism. To overcome these limitations, we introduce MORALISE, a comprehensive benchmark for evaluating the moral alignment of vision-language models (VLMs) using diverse, expert-verified real-world data. We begin by proposing a comprehensive taxonomy of 13 moral topics grounded in Turiel’s Domain Theory, spanning the personal, interpersonal, and societal moral domains encountered in everyday life. Built on this framework, we manually curate 2,481 high-quality image-text pairs, each annotated with two fine-grained labels: (1) topic annotation, identifying the violated moral topic(s), and (2) modality annotation, indicating whether the violation arises from the image or the text. For evaluation, we encompass two tasks, moral judgment and moral norm attribution, to assess models’ awareness of moral violations and their reasoning ability on morally salient content. Extensive experiments on 19 popular open- and closed-source VLMs show that MORALISE poses a significant challenge, revealing persistent moral limitations in current state-of-the-art models.} }
Endnote
%0 Conference Paper %T MORALISE: A Structured Benchmark for Moral Alignment in Visual Language Models %A Xiao Lin %A Zhining Liu %A Ze Yang %A Gaotang Li %A Ruizhong Qiu %A Shuke Wang %A Hui Liu %A Haotian Li %A Yuchen Yan %A Sumit Keswani %A Vishwa Pardeshi %A Huijun Zhao %A Wei Fan %A Hanghang Tong %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-lin26aa %I PMLR %P 74723--74744 %U https://proceedings.mlr.press/v306/lin26aa.html %V 306 %X Recently, vision-language models have demonstrated increasing influence in morally sensitive domains such as autonomous driving and medical analysis, owing to their powerful multimodal reasoning capabilities. As these models are deployed in high-stakes real-world applications, it is of paramount importance to ensure that their outputs align with human moral values and remain within moral boundaries. However, existing work on moral alignment either focuses solely on textual modalities or relies heavily on AI-generated images, leading to distributional biases and reduced realism. To overcome these limitations, we introduce MORALISE, a comprehensive benchmark for evaluating the moral alignment of vision-language models (VLMs) using diverse, expert-verified real-world data. We begin by proposing a comprehensive taxonomy of 13 moral topics grounded in Turiel’s Domain Theory, spanning the personal, interpersonal, and societal moral domains encountered in everyday life. Built on this framework, we manually curate 2,481 high-quality image-text pairs, each annotated with two fine-grained labels: (1) topic annotation, identifying the violated moral topic(s), and (2) modality annotation, indicating whether the violation arises from the image or the text. For evaluation, we encompass two tasks, moral judgment and moral norm attribution, to assess models’ awareness of moral violations and their reasoning ability on morally salient content. Extensive experiments on 19 popular open- and closed-source VLMs show that MORALISE poses a significant challenge, revealing persistent moral limitations in current state-of-the-art models.
APA
Lin, X., Liu, Z., Yang, Z., Li, G., Qiu, R., Wang, S., Liu, H., Li, H., Yan, Y., Keswani, S., Pardeshi, V., Zhao, H., Fan, W. & Tong, H.. (2026). MORALISE: A Structured Benchmark for Moral Alignment in Visual Language Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:74723-74744 Available from https://proceedings.mlr.press/v306/lin26aa.html.

Related Material