Synthetic measurements show that gold-standard performance degrades according to distinct functional forms when optimizing proxy reward models via RL or best-of-n, with coefficients scaling smoothly by reward model parameter count.
Observational overfitting in reinforcement learning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2022 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Scaling Laws for Reward Model Overoptimization
Synthetic measurements show that gold-standard performance degrades according to distinct functional forms when optimizing proxy reward models via RL or best-of-n, with coefficients scaling smoothly by reward model parameter count.