Ask an AI to generate a slide deck or dashboard as code, and you usually get something "correct but kind of sad-looking." Code-generation models can handle text and table content accurately, but since they're not writing while watching the canvas, their layout and color choices tend to fall flat. Image-generation models look great but misread text and numbers, and they only output flat, uneditable, unverifiable raster images. A paper from Yale University and Microsoft, "AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards," proposes combining the strengths of both. It feeds a code-generation model a reference image produced by an image generator from the same prompt, as aesthetic guidance, and optimizes seven reward channels, execution, text accuracy, table/chart structure, layout, and whitespace among them, each normalized independently via a structured "design graph" representation, using reinforcement learning (GDPO). On a 300-item benchmark, the 8B model's training alone lifted its overall score from 42.39 to 82.94 (+40.55), and the 32B model reached 85.89, beating GPT-5.5 (81.28) and Claude Opus 4.8 (80.39). It also leads frontier models on text accuracy (95.34%) and chart/table verification (90.37%), proving that looking good and being correct aren't a trade-off here. It's a genuinely practical escape from the usual choice between "looks great but the content is sketchy" and "the content is right but it looks sad." URL: #CodeGeneration# #DesignAI#