BRo-JEPA: AI World Model Achieves Zero-Shot Modular Arithmetic Generalization
A recent study presents BRo-JEPA, a world model capable of learning algebraic principles from visual data, as detailed in a publication on arXiv (identifier 2606.01372v2). The research investigates whether neural networks can grasp modular arithmetic through images of handwritten digits using the MNIST dataset and EMNIST letters. The authors evaluate traditional supervised baselines, which falter with unfamiliar operations, against BRo-JEPA, which encodes operations as rotations in latent space. This approach allows for zero-shot operation generalization without the need for fine-tuning. Results indicate that the top baseline achieves 54.54% zero-shot accuracy on MNIST and 25.13% on EMNIST, while BRo-JEPA outperforms significantly. This work enhances JEPA models and seeks to advance AI's ability in rule-based reasoning tasks.
Key facts
- Paper titled 'BRo-JEPA: Learning Modular Transformations in Latent Space' is available on arXiv with identifier 2606.01372v2.
- The announcement type is 'replace-cross' for version 2.
- BRo-JEPA is a JEPA-style world model that learns algebraic rules from visual inputs.
- The study uses MNIST and EMNIST letters as states and modular arithmetic operations as actions.
- Standard supervised and JEPA baselines with operation embeddings fail to extrapolate to unseen operations.
- BRo-JEPA represents arithmetic operations as rotations in latent space, capturing the cyclic structure of modular arithmetic.
- The best block-rotation supervised baseline achieves only 54.54% zero-shot accuracy on MNIST and 25.13% on EMNIST.
- BRo-JEPA enables strict zero-shot operation generalization.
Entities
—