<ul><li>A new approach for generalizable reward model in Reinforcement Learning from Human Feedback (RLHF) is proposed.</li><li>Existing reward models lack the ability to correctly evaluate unseen prompt-response pairs.</li><li>The proposed approach decomposes the reward value into prompt-free reward and prompt-related reward.</li><li>The new reward learning algorithm prioritizes data samples based on their prompt-free reward values.</li></ul>

Information-Theoretic Reward Decomposition for Generalizable RLHF

Discover more