HoneyBench - A benchmark for reward hacking in frontier models

HoneyBench is introduced as a new benchmark designed to evaluate reward hacking behaviors in advanced AI models. It aims to measure how frontier models might exploit reward functions in unintended ways, highlighting alignment challenges.