Summary of Mission Impossible: a Statistical Perspective on Jailbreaking Llms, by Jingtong Su et al.
Mission Impossible: A Statistical Perspective on Jailbreaking LLMs
by Jingtong Su, Julia Kempe, Karen Ullrich
First submitted to arxiv on: 2 Aug 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); 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 Large language models (LLMs) trained on vast amounts of text data with limited quality control can exhibit unintended or harmful behaviors, such as leaking information, spreading fake news or hate speech. To counter these issues, researchers have developed preference alignment techniques, fine-tuning the LLMs with crafted text examples of desired behavior. However, empirical evidence shows that even preference-aligned LLMs can be enticed to harmful behavior through adversarial prompt modifications. The paper provides theoretical insights into preference alignment and jailbreaking from a statistical perspective, highlighting the limitations of current approaches. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models (LLMs) are trained on lots of text data, but sometimes they say things we don’t want them to! To fix this, people try to teach LLMs good behavior by giving them examples. But it turns out that even when we do this, the LLMs can still be tricked into saying bad things. The authors of a new paper are trying to understand why this happens and how we can make LLMs safer without sacrificing their ability to help us. |
Keywords
» Artificial intelligence » Alignment » Fine tuning » Prompt
