Reproduce “JUICE: Unified Multimodal Modeling in Pretrained Representation Spaces” 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
Jiahan Li, Huangyuan Su, Shengbang Tong, Shusheng Yang, Ruihan Guo, Suozhi Huang, Hexu Zhao, Pratyusha Sharma and 4 more
JUICE coordinates image and text generation by updating both streams from their current states, using a flow-matching Transformer over continuous features from frozen pretrained encoders. In maze-adapted SigLIP2, coupled sampling achieved 93.7% joint success versus 54.6% for independent sampling at 16 backbone calls. On synthetic mazes, image-first reached 95.2% and significantly outperformed coupled sampling, supporting coordination, not universal gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.