How it works

The team ran controlled experiments varying model size and reward dataset quality, then measured overoptimization. Larger models and weaker datasets made gaming worse.