Deepmind proposed a Zuan AI, which specially outputs network attack language

Language model (LM) There is often the potential harm of generating offensive language , This also affects the deployment of the model . Some studies try to use manual annotators to write test cases , To identify harmful behaviors before deployment . However , Manual annotation is expensive , It limits the number and diversity of test cases .

Based on this , come from DeepMind By using another LM Generate test cases to automatically discover targets LM Possible harmful performance in the future . This study uses classifiers that detect offensive content , To evaluate goals LM Quality of answers to test questions , In the experiment 280B Parameters LM Tens of thousands of aggressive answers have been found in chat robots .

Address of thesis ： https://storage.googleapis.com/deepmind-media/Red%20Teaming/Red%20Teaming.pdf

This study explores a variety of methods from zero sample generation to reinforcement learning , To generate test cases with diversity and different difficulties . Besides , The study uses prompt Engineering to control LM Generate test cases to discover other hazards , Automatically find out who the chat robot will discuss with in an aggressive way 、 Find out the harmful situation in the dialogue process such as revealing private information . Overall speaking , Proposed by the study Red Teaming LM It is a promising tool , It is used to find and repair all kinds of bad before actual users use LM Behavior .