HoneyBench - A general benchmark for reward hacking in frontier models

HoneyBench is introduced as a comprehensive benchmark designed to evaluate reward hacking vulnerabilities in advanced AI models, aiming to improve their reliability and safety.