Reproduce “AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards” 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
Xin Gao, Yifan Yang, Xin Li, Pan Wang, Yan Li, Muzhao Tian, Mingxi Cheng, Bei Liu and 8 more
AesCode trains code models to create information-rich infographics, using prompts to determine content and generated images as non-authoritative design suggestions. It represents page elements and relationships in a design graph, then combines programmatic checks with graph-bound visual rewards. On a 300-instance benchmark, AesCode-32B scores 77.44 Visual versus 55.02 for Qwen3-VL-32B; Style remains below 60% for every model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.