WebGrader: Self-Evolving Programmatic Grader for Training LLMs in Web Development
A new research paper on arXiv (2608.06474) introduces WebGrader, a self-evolving programmatic grader designed to train large language models (LLMs) for web development tasks. The paper addresses the bottleneck of reward design in reinforcement learning, where hand-authored browser scripts are costly and VLM/GUI-agent graders may issue premature verdicts. WebGrader autonomously derives interaction flows from website requests, represents them as executable Flow Contracts, and uses execution outcomes as rewards. It materializes generated projects in a live browser, grounds actions against source code and DOM, and collects multimodal evidence. A residual-driven offline loop further refines the grader. The approach aims to improve the functional correctness of LLM-generated websites.
Key facts
- WebGrader is a self-evolving programmatic grader for training LLMs in web development.
- It addresses the reward design bottleneck in reinforcement learning for code generation.
- Hand-authored browser scripts are costly; VLM and GUI-agent graders may be premature.
- WebGrader derives interaction flows from website requests and represents them as Flow Contracts.
- Execution outcomes of Flow Contracts serve as RL rewards.
- It materializes projects in a live browser and grounds actions against source code and DOM.
- It collects visual, DOM, response, and persistent-state evidence.
- A residual-driven offline loop enhances the grader's performance.
- The paper is available on arXiv with ID 2608.06474.
Entities
Institutions
- arXiv