VF-Coder: Visual Feedback Enhances GUI Code Generation

Researchers have introduced VF-Coder, a multi-agent system that uses visual feedback to improve GUI code generation and debugging. Current LLM-based agents rely on text-output feedback, which fails for event-driven GUI programs with visual attributes. To address this, the team created InteractGUI Bench, a benchmark of 984 real-world desktop GUI tasks for evaluating interaction logic and visual structure. VF-Coder processes visual information to simulate user interactions and assess rendered interfaces, overcoming limitations of text-only approaches. The system targets the gap in automated GUI development, where existing methods cannot trigger GUI element logic or verify visual compliance with user needs. The work is detailed in arXiv:2604.19750.