Behavioral Task Sampling Improves Zero-Shot Offline RL

A recent study published on arXiv (2604.25496) introduces a technique aimed at enhancing zero-shot offline reinforcement learning by directly deriving task vectors from offline datasets instead of using random sampling. Traditionally, task-conditioned policies are trained with randomly chosen task vectors that establish linear reward functions across learned state representations. The researchers contend that this method results in less effective generalization. Their straightforward procedure for extracting reward functions can be incorporated into current algorithms, yielding an average improvement of 8% in zero-shot performance across various benchmark environments and baselines.