MobileForge: New Benchmark for Multi-Screen Mobile App Generation

Current benchmarks for design-to-code AI models primarily focus on generating single-page interfaces, which falls short of the complexity required for real-world mobile applications. Real apps demand multiple interconnected screens, shared components, and functional navigation within a buildable codebase. To address this gap, researchers have developed MobileForge, a novel benchmark specifically for project-level multi-screen mobile app generation. MobileForge comprises real mobile apps, human-reviewed screens, structured page-relationship annotations, and navigation test specifications. It enables a comprehensive five-axis evaluation covering build success, navigation, visual fidelity, code maintainability, and efficiency. Initial tests on six leading multimodal LLMs show that while models can generate compilable projects and reach correct pages, interactive navigation remains unreliable, and visual fidelity and code maintainability still require significant improvement. The benchmark also introduces state-isolated navigation testing and an anchor-referenced visual evaluation protocol to enhance reliability.