CLARIFY: Contrastive Preference Reinforcement Learning for Untangling Ambiguous Queries

In preference-based reinforcement learning (PbRL), humans find it hard to compare behaviorally similar trajectory segments, making it difficult to provide clear preference signals and hurting label efficiency. CLARIFY addresses this by designing a contrastive-learning objective in the trajectory embedding space that selects pairs of segments humans can compare, thereby improving label efficiency.

Links: arXiv · Code

Paper | Code