BRo-JEPA: Learning Modular Arithmetic in Latent Space

arXiv 2026 56.3 method

TLDR

A JEPA-style latent world model with block-rotation predictor learns modular arithmetic rules, achieving 99.46% zero-shot accuracy on MNIST digits.

Reasoning

The paper presents a novel architecture that enforces circular structure in latent space, enabling strong generalization on a synthetic task. However, it lacks real-world evaluation and only tests on a single, simple domain (modulo-10 arithmetic with MNIST).

Read-first score

Read-first score 56.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 22.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

92.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.92165707

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,github

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,result

Topical relevance 29%

31.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 465.

Keyword Scores

world model

world dynamics prediction

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Block-rotation predictor that imposes the circular structure of modulo-10 arithmetic in latent space
JEPA-style latent world model for learning modular arithmetic with MNIST digits as states and operations as actions
Achieving strong zero-shot generalization to unseen modular arithmetic operations

Methodology

The paper uses MNIST digits as states and modular arithmetic operations as actions in a JEPA-style latent world model. It compares standard supervised baselines and JEPA models with additive operation embeddings against a proposed block-rotation predictor that enforces the circular structure of modulo-10 arithmetic in latent space. The best model is a ResNet-based JEPA block-rotation model evaluated on zero-shot and rollout accuracy.

Key Results

The best ResNet-based JEPA block-rotation model achieves 99.46% zero-shot accuracy and 99.46% rollout accuracy on unseen modular arithmetic operations.

Resources

Paper Code