WebGrader Trains LLMs for Web Development with Self-Evolving Grader.

Training Large Language Models (LLMs) to generate functional websites from natural language descriptions is challenging, primarily due to the difficulty of designing effective reward signals for reinforcement learning. Existing methods either require costly manual script writing or rely on visual/GUI agents that might misjudge functionality. WebGrader introduces a novel solution: a self-evolving programmatic grader. WebGrader autonomously derives required interaction flows from web requests, representing them as executable "Flow Contracts." It then materializes the generated website in a live browser, grounding actions against the source code and DOM, and collecting comprehensive evidence. A key innovation is its residual-driven offline loop, which discovers and validates reusable verifier skills, freezing them into a skill graph before policy training. This ensures that a "Pass" verdict is issued only after observing the requested state transition, providing a highly accurate reward signal. WebGrader has shown significant improvements, training an 8B policy to a 52.01% functional success rate on WebGen-Bench, outperforming other reward systems and even larger models.