ASTPrompter: Weakly Supervised Automated Language Model Red-Teaming to Identify Low-Perplexity Toxic Prompts

원문 초록

Conventional approaches for the automated red-teaming of large language models (LLMs) aim to identify prompts that elicit toxic outputs from a frozen language model (the defender). This often results in the prompting model (the adversary) producing text that is unlikely to arise during autoregression. In response, we propose a reinforcement learning formulation of LLM red-teaming designed to discover prompts that both (1) elicit toxic outputs from a defender and (2) have low perplexity as scored by that defender. These prompts are the most pertinent in a red-teaming setting because the defender generates them with high probability. We solve this formulation with an online and weakly supervised form of Identity Preference Optimization (IPO), attacking models ranging from 137M to 7.8B parameters. Our policy performs competitively, producing prompts that induce defender toxicity at a rate of 2-23 times higher than baseline across model scales. Importantly, these prompts have lower perplexity than both automatically generated and human-written attacks. Furthermore, our method creates black-box attacks with 5.4-14 times increased toxicity. To assess the downstream utility of our method, we use rollouts from our policy as negative examples for downstream toxicity tuning and demonstrate improved safety.