Large language model (LLM) agents select actions through next-token prediction or chain-of-thought reasoning, without an explicit model of how those actions transform the world. We ask whether the Joint Embedding Predictive Architecture (JEPA), originally proposed for self-supervised visual representation learning, can serve as the foundation for a reward-free world model in the text-based agent setting. We present Agentic-JEPA, a system that trains a lightweight transformer predictor to map (state, action) pairs to next-state embeddings, using decoder-only LLM backbones as state and action encoders and an EMA target encoder to prevent representation collapse. At inference time, the agent plans without any reward signal by scoring candidate actions via cosine similarity between predicted outcomes and a goal embedding, with optional k-step lookahead. On a controlled benchmark of 17 text-based environments spanning three transition dynamics (linear, branching, stochastic) and a strict 10 train / 7 held-out test split, our experiments reveal a sharp generalization boundary. Agentic-JEPA achieves 100% success and near-oracle planning efficiency on all in-distribution environments, demonstrating for the first time that JEPA can learn text-environment dynamics from self-supervised trajectories alone, yet fails completely (0% success) on every out-of-distribution environment, regardless of dynamics type. We formalize two necessary conditions for cross-environment transfer (encoder coverage and structural alignment) and show, including through a t-SNE visualization of the learned latent space, that neither holds: the encoder partitions representations by environment identity rather than by functional role. Two secondary findings round out the picture: k-step lookahead degrades performance (100% to 40% at k=3) due to compounding prediction error, and scaling the backbone from GPT-2 to a 4-bit quantized Qwen 2.5-1.5B does not recover OOD success. These findings characterize both the promise and the current boundaries of self-supervised world models for text-based agents, and turn the 0% OOD result from a flat negative into a diagnostic with concrete, testable targets for future work. Code, environments, and training trajectories are publicly available at https://github.com/franfj/agentic-jepa.
Francisco-Javier Rodrigo-Ginés (Sat,) studied this question.