Zero-shot reinforcement learning aims to let an agent handle new tasks without additional training. Successor-feature-based approaches typically pre-train a task encoder and a successor-feature predictor on offline data that contains no rewards, then recover a policy at test time through a closed-form linear regression.

The new method, Sparse-TD-JEPA, extends this framework by introducing sparse latent task representations. The sparsity is intended to make the learned task representations more efficient and generalizable, which should improve zero-shot performance when the policy is solved via the linear regression step.

The abstract describes the overall architecture and training setup but does not include experimental results or comparisons. As such, the reported improvement is presented as the method's design goal rather than a demonstrated outcome in the available text.