Show HN: Lar-JEPA – A Testbed for Orchestrating Predictive World Models (github.com)

Lar-JEPA introduces a revolutionary framework for orchestrating predictive world models, breaking away from conventional prompt-based approaches used in autoregressive large language models (LLMs). Traditional models, which rely on a linear sequence of text, often fail when encountering hallucinations or poorly reasoned actions, jeopardizing entire execution plans. In contrast, the Lar-JEPA architecture utilizes a Joint Embedding Predictive Architecture (JEPA) capable of predicting abstract environmental states through mathematical representations, eliminating the risks tied to text-based decision-making. The Lar framework, designed for non-LLM predictive agents, features a deterministic execution spine and a cognitive memory layer. It allows agents to seamlessly process high-dimensional latent tensors and reroute actions before they materialize, enhancing safety and efficiency. Key technical advancements include native tensor logging for real-time feedback, continuous learning through a Default Mode Network (DMN) that consolidates experiences, and mathematical routing for decision-making without the need for textual prompts. This innovation not only enhances the capabilities of AI agents but paves the way for a new generation of cognitive architectures that prioritize planning and environmental interaction over traditional text outputs.