Scaling Laws for Reward Model Overoptimization

By Leo Gao, John Schulman, Jacob Hilton. 2022

1,126 citations recorded by the source corpus.

In reinforcement learning from human feedback, it is common to optimize against a reward model trained to predict human preferences. Because the reward model is an imperfect proxy, optimizing its value too much can hinder ground truth performance, in accordance with Goodhart's law. This effect has been frequently observed, but not carefully measured due to the expense of collecting human preference data. In this work, we use a synthetic setup in which a fixed"gold-standard"reward model plays the role of humans, providing labels used to train a proxy reward model. We study how the gold reward model score changes as we optimize against the proxy reward model using either reinforcement learning or best-of-$n$ sampling. We find that this relationship follows a different functional form depending on the method of optimization, and that in both cases its coefficients scale smoothly with the number of reward model parameters. We also study the effect on this relationship of the size of the reward model dataset, the number of reward model and policy parameters, and the coefficient of the KL penalty added to the reward in the reinforcement learning setup. We explore the implications of these empirical results for theoretical considerations in AI alignment.

Subfields: Artificial Intelligence, Alignment & RLHF

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

  1. Mohit Bansal
  2. Sergey Levine
  3. Yoshua Bengio
  4. Percy Liang
  5. Luke S. Zettlemoyer
  6. Jianfeng Gao
  7. Quoc V. Le
  8. Kyunghyun Cho

In Alignment & RLHF

  1. Angelica Chen
  2. Sergey Levine
  3. Jan Leike
  4. Tomasz Korbak
  5. Ethan Perez
  6. Anca Dragan
  7. Rafael Rafailov
  8. Jeff Wu

Citations in the corpus