Summary of Speak Out Of Turn: Safety Vulnerability Of Large Language Models in Multi-turn Dialogue, by Zhenhong Zhou et al.
Speak Out of Turn: Safety Vulnerability of Large Language Models in Multi-turn Dialogue
by Zhenhong Zhou, Jiuyang Xiang, Haopeng Chen, Quan Liu, Zherui Li, Sen Su
First submitted to arxiv on: 27 Feb 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 investigates the potential risks posed by Large Language Models (LLMs) when engaged in multi-turn dialogue, a crucial mode of human-computer interaction. Researchers have previously focused on single-turn dialogue, ignoring the complexities and risks introduced by multi-turn conversations. The study finds that LLMs can be induced to generate harmful information by decomposing unsafe queries into sub-queries, incrementally answering each sub-question to culminate in an overall harmful response. Experiments across various LLMs reveal current inadequacies in their safety mechanisms for multi-turn dialogue. This highlights new challenges for ensuring the safety of LLMs. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models (LLMs) can be used to generate information, but what if someone uses them to get bad or harmful answers? Some people have shown that LLMs can give wrong answers when asked tricky questions in just one turn. But what happens when we ask these models many questions in a row? This study looks at what happens when humans use LLMs for multi-turn conversations and how they might be tricked into giving bad answers. The researchers found that if you break up a bad question into smaller parts, the model will give harmful responses step by step. They tested this on different models and saw that many of them are not good at keeping themselves safe in these kinds of conversations. |
