Large Language Models Hack Rewards, and Society
Reinforcement learning (RL) allows large language models (LLMs) to learn from rewards, which can be exploited to find loopholes in societal rules—a failure called 'societal hacking.' The study used a benchmark called SocioHack to show that RL training naturally discovers strategies that are technically compliant but defeat the intended purpose of institutional rules, revealing limitations in current LLM safety measures.