Title: BRo-JEPA: Mastering Modular Arithmetic Within Latent Representations

Abstract: Do artificial neural networks possess the capacity to internalize abstract algebraic principles, or are they limited to rote memorization of training examples? To address this question, we employ MNIST digits to represent states and modular arithmetic operations as actions within a JEPA-style latent world model framework. While conventional supervised baselines and JEPA architectures equipped with additive operation embeddings successfully model observed operations, they lack the robustness to generalize reliably to novel scenarios. To overcome this limitation, we propose a block-rotation predictor that encodes the cyclic nature of modulo-10 arithmetic directly into the latent space. This architectural innovation facilitates exceptional zero-shot generalization capabilities. Notably, our top-performing ResNet-based JEPA model utilizing block-rotation achieved a zero-shot accuracy of 99.46% and a rollout accuracy of 99.46%. These findings indicate that latent world models are capable of acquiring symbolic transformation rules, provided that the model architecture aligns with the underlying structural properties of the task. The source code for this study is available at https://github.com/DL-World-Models/mnist-math.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC