Detecting Reward Hacking When Benchmark Graders Become Reward Models

Microsoft's approach of repurposing benchmark graders as reinforcement-learning reward models triples enterprise agent reliability, but it also raises reward-hacking and generalization risks that teams must monitor.