Agentic-JEPA: A Self-Supervised World Model for Planning in Text-Based Agent Environments

Abstract

We propose Agentic-JEPA, a self-supervised world model that combines the Joint Embedding Predictive Architecture (JEPA) with agentic planning capabilities for text-based environments. The model learns latent representations of environment states and predicts future states in embedding space, enabling efficient planning without pixel-level reconstruction.