Reward and Restraint

Optimize a three-answer policy and discover why more reward can mean less truth.

From Chapter 10: From Base Model to Assistant