WebGrader: Training LLMs for Web Development with Self-Evolving Programmatic Grader
- Source
- arxiv.org
- Author
- Boshui Chen, Huiping Liu, Shaolei Zhang
- Date
Context
WebGrader rewards a website-building model only after observing the requested behavior in a browser. It turns a website request into executable interaction flows, locates the relevant controls using both source code and the live page, and gathers evidence of the resulting state. This addresses a grading problem: a page can look convincing before anyone establishes that its requested interactions work. Research abstract
The system develops reusable verification skills offline and freezes them before model training. On WebGen-Bench, the authors report 52.01% functional success for an eight-billion-parameter model trained with WebGrader, 7.88 percentage points above a matched reward based on appearance and scripts. The result connects the choice of training feedback to better measured website functionality. Reported evaluation
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
articleReinforcement Learning for Design Taste — Jasper Lujasperlu.com
articleSERPO: Self-Evolving Rubric Policy Optimization for Open-Ended Test-Time Reinforcement LearningJianze Wang, Kunwang Zheng, Ying Liu, Yu Cao, Qilong Zhang, Jinlong Chen, Hua Yang, Qianglong Chen
articleInducing Reward-Free Judging Rubrics that Reduce Over-Crediting in Agent EvaluationDarragh Quinn, David Dylan, Roisin Healy, Fionn Carroll, Maeve Donnelly, Cormac Sheehan
Checking sign-in…
Loading comments…