Playground 4 — Reward Hacking Simulator

Drag the training steps slider. Watch the RM score climb while true output quality peaks and then degrades — the signature of reward hacking.

Training steps 0

📏 Response Length

avg tokens120

RM spuriously rewards verbosity → padding

🎭 Sycophancy Score

agreement rate12%

Model learns to agree rather than inform

⭐ Reward Model Score

RM score3.2

Rises — increasingly from exploits

✅ True Human Quality

actual quality3.1

Improves, then plateaus, then degrades

Move the slider to begin simulated RL training and observe reward hacking emerge over time.