0. Updates & motivations

0.1. Updates

0.2. Motivations

• My motivations of writing this survey are folded as follows.

1. Introduction

Based on my unfinished survey on the topic LLM safety, I am motivated to write down some summaries and thoughts on this both old and infant topic.

• Why old?
• Why infant?

My logic of structuring this article is as follows:

1. Firstly (Sec. 2), I want to introduce a few works about toxicity and truthfulness issue of LLMs. They mark a start of the general field of LLM safety.
2. Secondly (Sec. 3), I try to describe the extension and intension of LLM safety as its definition.
3. Then (Sec. 4), I discuss a number of papers that attempt to explain and understand the origins of LLM unsafety from statistical, mechanistic, and data-centric perspectives.
4. Fourthly (Sec. 5), I summarize and exemplify several jailbreaking methods.
5. Lastly (Sec. 6), I list and discuss several directions on how to defend.
6. Bonus (Sec. 7), where I discuss from inside out the path towards LLM safety based on my understanding.

2. Toxic and untruthful, the start

• Surveyed papers

This section motivates LLM safety by reflecting on early works that study toxicity and untruthfulness problem with early LLMs like GPT-3. I think LLM safety, as a growing interest for both research community and industry, started from a few researchers who realized and advocated that: