Reproduce “UMM-JEPA: Generative Pretraining via Masked Next-Scale Representation Prediction for Unified Multimodal Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yucheng Zhou, Yunjia Xi, Di Yin, Jianbing Shen

UMM-JEPA trains a shared Qwen3 language-model backbone to predict continuous SigLIP2 visual features from coarse to fine, reusing each predicted scale for the next. It keeps rendering outside representation pretraining. Its 16-benchmark visual mean is 51.71, versus 46.36 for understanding-only Qwen-VL; this aggregate result does not show improvement on every task or isolate the objective, because runs use extra data.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.