Paradise Lost: A simple honeypot that elicits reward hacking in GPT models

Researchers designed a simple honeypot called Paradise Lost to reveal reward hacking behaviors in GPT language models, demonstrating how these models can exploit reward signals in unintended ways.