Short Answer

Stabilizing an Underdetermined Reward Model

A machine learning engineer observes that their reward model, trained to prefer one text completion over another, is learning the correct relative rankings. However, the absolute values of the rewards assigned to the completions are growing excessively large during training, leading to instability. Explain why this phenomenon is a classic sign of an underdetermined problem in this context, and describe how adding a penalty for large reward values to the training objective helps to mitigate this issue.

0

1

Updated 2025-10-10

Contributors are:

Who are from:

Tags

Ch.4 Alignment - Foundations of Large Language Models

Foundations of Large Language Models

Foundations of Large Language Models Course

Computing Sciences

Application in Bloom's Taxonomy

Cognitive Psychology

Psychology

Social Science

Empirical Science

Science