Summary of Safe Inputs but Unsafe Output: Benchmarking Cross-modality Safety Alignment Of Large Vision-language Model, by Siyin Wang et al.
Safe Inputs but Unsafe Output: Benchmarking Cross-modality Safety Alignment of Large Vision-Language Model
by Siyin Wang, Xingsong Ye, Qinyuan Cheng, Junwen Duan, Shimin Li, Jinlan Fu, Xipeng Qiu, Xuanjing Huang
First submitted to arxiv on: 21 Jun 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary In this paper, researchers address a critical concern in the development of Artificial General Intelligence (AGI): ensuring its safety and ethical alignment. To evaluate the cross-modality safety alignment of AGI systems, they introduce the Safe Inputs but Unsafe Output (SIUO) challenge, which simulates scenarios where individual modalities are safe but could lead to unsafe outputs when combined. The authors develop a benchmark, SIUO, covering 9 critical safety domains, and test it on popular language models like GPT-4V and LLaVA. Surprisingly, their findings reveal significant safety vulnerabilities in these models, highlighting the need for more advanced AI systems that can reliably interpret and respond to complex real-world scenarios. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper aims to make AGI safer by creating a new challenge called SIUO. This challenge looks at how different parts of an AI system work together and if they could cause problems even if each part works fine on its own. The authors created a test for this challenge with 9 examples of safety issues, like self-harm or privacy violations. They then tested popular language models to see how well they do on the SIUO challenge. What they found was that these models have some big problems when it comes to staying safe and not causing harm. |
Keywords
» Artificial intelligence » Alignment » Gpt
