WebCoder trains frontend agents on a living project: 72.8 requirement satisfaction beats Claude-Opus-4.8's 65.1
Synopsis
WebCoder-27B, trained during an internship at Zhipu AI, keeps reproduction, generation, modification and repair on one evolving project state and scores reinforcement learning by how well the executable application meets task requirements. On the new Webdev-Bench its requirement satisfaction rises from 64.8 for the Qwen3.8-27B base model to 72.8, above Claude-Opus-4.8 at 65.1; on Vision2Web its average visual and functional score rises from 41.1 to 62.4, on par with Claude-Opus-4.8 at 62.5.
Interpretation
Frontend agents are no longer treated as one-shot screenshot-to-code tools; a single policy must reproduce, generate, modify and repair on the same evolving application, and WebCoder-27B is trained for that composition. Earlier work trained and evaluated design reproduction, refinement from rendered feedback and interactive site generation separately, while real requests need those abilities to share one codebase and one application state. On Webdev-Bench, 40 composite tasks, WebCoder-27B reaches a requirement satisfaction score of 72.8 against 64.8 for the Qwen3.8-27B base model and 65.1 for Claude-Opus-4.8, with the other proprietary models between 48.0 and 59.6, and it delivers all 40 tasks.
The training signal comes from whether the final application satisfies the task requirements, not from what the agent says it did. An independent verifier scores only evidence from building, running and inspecting the application, and the quality score is gated by build, entry-point resolution and browser rendering. On the 66 interactive-frontend tasks of Vision2Web, WebCoder-27B reaches an average visual and functional score of 62.4 against 41.1 for the base model and 62.5 for Claude-Opus-4.8, and its delivery rate rises from 48.5% to 80.3%.
Intermediate progress is estimated from the same requirement checklist, so the final reward can be spread across development stages. A terminal reward alone says whether the task finished but not which edits advanced it or introduced regressions, and the effect of an edit only becomes visible after building, running and inspecting. Ablations on Vision2Web L2 lower the average from 62.4 to 52.6 without artifact-grounded credit assignment and to 54.9 without context compaction, and both variants stay above the base model's 41.1.
Long trajectories cross a single context window through context compaction and message-tree sample construction. Development history outgrows the model window, and training on compacted history repeatedly would count the same generated content in the loss more than once. Training uses a 116,736-token rollout context and a 65,536-token training context, and ownership masks keyed by message position and content let shared generated history contribute to the policy loss at most once.
Perspective
Evaluation covers the 66 interactive-frontend tasks of Vision2Web and the 40 composite tasks of Webdev-Bench, with every model using the same agent framework, tools, prompts, container environment, time budget and judges; training tasks are built from synthetic product specifications and screenshots of public web pages, and agents run in isolated containers with no access to external services beyond package registries.
Webdev-Bench terminal acceptance is performed by a language-model agent whose repeated-grading consistency is measured separately in the appendix; case studies also include a task that fails acceptance, where a floating-point remainder in a price quote leaves one requirement unsatisfied. The authors state that the model, benchmark and training and evaluation code will be open-sourced upon acceptance.