Red Teaming LLMs with LLMs: the DeepMind paper on automated red teaming
In one sentence Perez et al. (DeepMind) use a language model to automatically generate test cases against a 280B-parameter chatbot, uncovering tens of thousands of offensive replies.
Traditional red teaming relies on people hand-writing questions designed to make a system fail. It is expensive, and it limits how many and how varied the tests can be.
The paper by Ethan Perez and colleagues at DeepMind proposes having another language model, the "red LM", write the tests. The target model's replies are then judged by a classifier trained to detect offensive content.
Applied to a 280-billion-parameter chatbot, the method finds tens of thousands of offensive replies. The authors present it as one promising tool, among many needed, for finding undesirable behaviors before they reach users.
Companies
DeepMind, New York University
Tools
Gopher
Tags
Sources