A Systematic Investigation of RL-Jailbreaking in LLMs
Researchers have conducted a systematic investigation into Reinforcement Learning (RL) jailbreaking techniques used against large language models (LLMs). Their analysis deconstructs the RL framework, examining aspects like reward functions, action spaces, and episode lengths to understand why these methods are effective. The study found that RL jailbreakers successfully compromised targeted models and safeguards, with environment formalization, particularly dense rewards and extended episode lengths, being the primary drivers of success. AI
IMPACT Identifies key factors in RL jailbreaking, offering insights for developing more robust LLM defenses.