WebGrader trains LLMs to build websites with 52% success

Researchers from institutions including Boshui Chen, Huiping Liu, and Shaolei Zhang introduced **WebGrader**, a self-evolving programmatic grader designed to train large language models (LLMs) for web development tasks. The tool addresses a critical bottleneck in reinforcement learning for LLMs: reward design. Traditional methods rely on expensive hand-authored browser scripts or scalable but potentially premature visual/GUI-agent evaluations. WebGrader instead dynamically derives interaction flows from natural-language prompts, executes them in a live browser environment, and validates outcomes against executable Flow Contracts—ensuring verdicts are only issued after observing decisive states.

On the **WebGen-Bench** benchmark, WebGrader trained an 8B-parameter policy to a **52.01% functional success rate**, surpassing a matched script-based reward by **7.88 points** and outperforming commercial models like o4-mini and DeepSeek-v4-flash. On the **WG-core-250** subset, the policy achieved a **perfect score of 44.953**, exceeding even Qwen3-Coder-480B. The system’s innovation lies in separating test planning, action grounding, evidence collection (visual/DOM/response/state), and semantic judgment, enabling precise validation of user-requested transitions.