Gary Marcus

Rank 13 of 47
|
Score 83
2/13/2025, 6:39:55 PM
X
In reply to:
Paul Röttger
@paul_rottger
·
364d
All of MSTS is permissively licensed and available now. Check out the MSTS preprint for more details, or go to GitHub / Hugging Face to access the dataset 👇

https://arxiv.org/abs/2501.10057
https://github.com/paul-rottger/…
https://huggingface.co/datasets/…
Paul Röttger
@paul_rottger
·
364d
MSTS was a huge group effort by 22 co-authors. Special thanks to @bertievidgen for co-leading the project. To core contributors @peppeatta @felix_friedri @jagoldz @AliciaVParrish. And also to @MLCommons for making this project possible in the first place!
Paul Röttger
@paul_rottger
·
364d
Throughout the paper, we used manual annotation to produce maximally reliable results. But can we automate safety assessments?

We find that even the best commercial VLMs struggle with this task.
Paul Röttger
@paul_rottger
·
364d
To isolate the impact of multimodality on safety, we created text-only versions of all our multimodal test prompts.

We find that VLMs produce unsafe responses for at least some multimodal prompts *only because they are multimodal*.
Paul Röttger
@paul_rottger
·
364d
We also manually translated MSTS into 10 other languages.

For the multilingual open VLM that we test, we find clear safety differences across languages. Notably, MiniCPM was developed in China and indeed produces the least unsafe responses in Chinese.
Paul Röttger
@paul_rottger
·
364d
When tested with MSTS, several open VLMs show clear safety issues. Not only do some VLMs give explicitly unsafe responses, but many are *safe by accident*, misunderstanding a high % of test prompts.

Commercial models like Claude and GPT, on the other hand, do quite well on MSTS.

The statement is a supportive acknowledgment of related work in the field of multimodal safety for vision-language models. It highlights collaboration and shared research interests.

  1. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding and collaboration by acknowledging similar work done by another researcher, fostering a sense of community and shared goals in the research field. [+1]