WebCoder trains one agent to handle changing frontend projects
Questa pagina non ha ancora una traduzione verificata. È mostrata la versione inglese. Leggi la versione inglese
In parole semplici
WebCoder is a frontend development agent trained to work on one runnable project as it changes. It can edit code, run the application, and inspect the result, so one agent can handle design copying, feature building, changes, and repairs together.
The authors trained it using checks of whether project requirements were met, plus estimates of progress between stages. On 40 tasks with requests revealed over time, it scored higher than its starting model on requirement satisfaction. This is evidence on those tasks, not every kind of frontend work.
La ricerca
When a frontend project gains requirements over several rounds, can an agent keep earlier behavior working while adding each new feature? A useful answer depends on evaluating the whole evolving application, because a locally successful change may still break something already built.
WebCoder trains one frontend agent in a persistent runnable project, using final requirement checks and progress between stages to guide learning.
The authors trained WebCoder in a shared environment where it could edit, run, and inspect one persistent project. An independent verifier checked the final application against task requirements; estimates of requirement progress between development segments helped assign training credit.
On Webdev-Bench, 40 tasks reveal requests in rounds and then recheck earlier requirements. WebCoder scored 72.8 on requirement satisfaction, against 64.8 for its Qwen3.8-27B base model; this score averages, across tasks, the share of requirements whose checklist items all received the top grade. That supports its ability to maintain combined requirements in this test.
The results cover 40 composite Webdev-Bench tasks and 66 titles at Vision2Web’s Interactive Frontend level. They do not establish performance on other task collections or other Vision2Web levels.
Analisi
Dettagli
- Autori del paper originale
- Zijun Dou, Zhen Yang, Xijun Liu, Dongzhuoran Zhou, Aohan Zeng
- Inviato ad arXiv
- Licenza
- CC BY 4.0