Summary of Cosafe: Evaluating Large Language Model Safety in Multi-turn Dialogue Coreference, by Erxin Yu et al.
CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference
by Erxin Yu, Jing Li, Ming Liao, Siqi Wang, Zuchen Gao, Fei Mi, Lanqing Hong
First submitted to arxiv on: 25 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 This paper addresses a critical concern in the development of large language models (LLMs): ensuring their safety from malicious attacks. Previous research has primarily focused on single prompt attacks or goal hijacking, but this study explores LLM safety in multi-turn dialogue coreference. The authors create a dataset of 1,400 questions across 14 categories, each featuring multi-turn coreference safety attacks. They then evaluate five widely used open-source LLMs under these attack scenarios, finding that the highest attack success rate is 56% with the LLaMA2-Chat-7b model and the lowest at 13.9% with the Mistral-7B-Instruct model. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks into making sure big language models are safe from being tricked or attacked. Right now, there’s no one doing this kind of research for multi-turn conversations. The researchers make a special dataset with questions and attacks that test how well these language models can keep their safety. They try out five different kinds of language models on this test and find out which ones are most vulnerable to being tricked. |
Keywords
» Artificial intelligence » Coreference » Prompt
