“In experiments across four RL environments, models with higher capability (larger size, higher action resolution, better observation fidelity, more training) tend to achieve higher proxy rewards but decreased true rewards.”