WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic Reinforcement Learning

Panel (a) shows frontend reproduction, generation and modification studied as separate task categories. Panel (b) shows one user request composing those capabilities over a continuous, stateful development process with multi-turn multimodal refinement. — **From isolated frontend tasks to general frontend development.** (a) Existing work typically studies frontend reproduction, generation, and modification as separate task categories. (b) In real-world development, a single user request often requires multiple capabilities to be composed within a continuous, stateful, and iterative development process, followed by multi-turn and multimodal refinement over the evolving application state.

Video

WebCoder in two minutes. Every number is taken from the paper. The editor, terminal and browser footage come from a real Webdev-Bench deliverable: the agent's own source, rebuilt and driven in a browser. The Vision2Web pages are the official reference screenshots and the agent's rendered output. The credit-assignment panel is an illustration of the mechanism, and is labelled as such. Music: “Inspired” by Kevin MacLeod (incompetech.com), licensed under CC BY 4.0.

Abstract

Real-world frontend development requires agents to compose reproduction, generation, modification, and repair over an evolving application state. Learning a general policy for this process requires translating heterogeneous execution feedback into task-aligned rewards and effective credit assignment across long sequences of interdependent actions.

We introduce WebCoder, a general frontend development agent trained with artifact-grounded agentic reinforcement learning. Our key idea is to assess diverse development activities through their effects on a shared executable artifact and its satisfaction of task requirements. We build Webdev-World, an executable environment that integrates code editing, application execution, browser inspection, and on-demand skills while preserving project state across interactions. An independent verifier evaluates the final application against task requirements to provide terminal rewards, while changes in intermediate artifact progress redistribute the resulting rollout advantage across development segments. Context compaction and message-tree sample construction support training on trajectories beyond a single context window while avoiding repeated loss computation on shared generated history.

We further introduce Webdev-Bench to evaluate composed development requirements within a single project. On Webdev-Bench, WebCoder-27B raises requirement satisfaction from 64.8 for its Qwen3.8-27B base model to 72.8, exceeding Claude-Opus-4.8 (65.1) and GPT-5.4 (59.6); on Vision2Web, it improves the average visual and functional score from 41.1 to 62.4, on par with Claude-Opus-4.8. Ablation studies confirm the contributions of artifact-grounded credit assignment and long-horizon training.

Method

Webdev-World

An executable environment that integrates code editing, application execution, browser inspection, and on-demand skills, while preserving project state across interactions. The agent sees the consequences of its own edits the way a developer does, through build results and the rendered page.

Artifact-grounded reward and credit assignment

An independent verifier scores the final application against the task's requirement checklist to produce the terminal reward. Changes in intermediate artifact progress then redistribute that rollout advantage across development segments, so a segment that actually moved the application forward is credited for it, while the token-weighted mean advantage set by the terminal reward is preserved.

Long-horizon training

Context compaction keeps long development sessions inside the context window, and message-tree sample construction lets a single trajectory train as several samples without computing the loss twice on shared generated history.

Webdev-Bench

Existing frontend benchmarks score one capability at a time. Webdev-Bench instead evaluates composed development requirements inside a single project: 40 tasks whose requirements arrive over several rounds, each with a hidden checklist, graded against the running application rather than against the source code.

Task statistics for Webdev-Bench, covering task tiers, starting conditions and requirement counts. — **Task statistics.** Tiers, starting conditions, and the distribution of requirements per task.

Results

**Main results on Vision2Web and Webdev-Bench.** Vision2Web (L2-frontend, 66 titles) reports the visual score (VS), functional score (FS), their average (Avg), and the delivery success rate (DSR). Webdev-Bench (40 tasks) reports requirement satisfaction (RS) together with its visual (VS), functional (FS), and interaction (IS) components, their average (Avg), and DSR. Undelivered tasks count as 0 in all score columns.
Model	Vision2Web				Webdev-Bench
Model	VS	FS	Avg	DSR	RS	VS	FS	IS	Avg	DSR
Proprietary models
Claude-Opus-4.8	59.1	65.9	62.5	98.5%	65.1	75.0	70.7	84.1	73.7	100.0%
GPT-5.4	51.3	64.4	57.9	97.0%	59.6	72.7	70.0	81.0	70.8	100.0%
GPT-5	23.3	64.7	44.0	98.5%	53.5	64.0	65.8	74.6	64.5	100.0%
Gemini-3-Pro	13.3	14.6	14.0	65.2%	48.0	65.8	58.5	68.5	60.2	100.0%
Gemini-3-Flash	13.1	27.9	20.5	78.8%	56.6	65.8	64.8	85.0	68.1	95.0%
Seed-1.8	7.0	13.3	10.2	36.4%	52.6	66.5	63.8	71.5	63.6	100.0%
Open-source models
Kimi-K2.5	35.1	61.0	48.1	92.4%	46.8	61.5	55.0	67.7	57.7	82.5%
MiMo-V2.5	38.2	66.5	52.4	87.9%	14.9	25.2	21.2	22.6	21.0	65.0%
Step-3.7-Flash	22.1	47.3	34.7	95.5%	10.8	20.0	14.0	19.0	15.9	60.0%
Qwen3.8-27B (our base model)	36.4	45.7	41.1	48.5%	64.8	77.9	68.8	79.8	72.8	95.0%
Qwen3.5-397B-A17B	35.6	67.1	51.4	100.0%	28.7	39.0	31.9	36.5	34.0	80.0%
Qwen3.5-122B-A10B	20.2	39.0	29.6	87.9%	24.6	30.4	28.9	33.1	29.3	72.5%
Qwen3-VL-32B-Instruct	0.0	0.0	0.0	0.0%	2.5	6.3	5.4	4.4	4.6	50.0%
Qwen3-VL-8B-Instruct	0.0	0.0	0.0	0.0%	0.0	0.8	0.0	0.0	0.2	30.0%
WebCoder-27B	57.7	67.1	62.4	80.3%	72.8	81.9	73.8	90.8	79.8	100.0%

Bold marks the best value in a column across all models; underline marks the second best.

Examples

WebCoder on Vision2Web caplugs and Webdev-Bench task c22: the reference homepage next to WebCoder's render, and the c22 application state after each requirement round. — **WebCoder on Vision2Web (caplugs) and Webdev-Bench (c22).** Left: the reference homepage of caplugs and WebCoder's render. Right: c22 states after each round, each cropped to the region the requirement changes. R2 tightens the hero and removes the investor row; the first attempt at R3 keeps the five “Why SEDA” parts in the old staggered layout and is rejected by the gate, while the second attempt compacts them into two rows and passes; R4 adds a filterable newsroom below the hero.

Round-by-round requirements and the resulting application states for Webdev-Bench task c22. — **A Webdev-Bench session.** Requirements arrive round by round over one evolving project.

Gallery of Vision2Web reproductions produced by WebCoder next to their reference screenshots. — **Vision2Web gallery.** Reference screenshots and WebCoder's reproductions.

Gallery of final deliverables produced by WebCoder across Webdev-Bench tasks. — **Webdev-Bench gallery.** Final deliverables across tasks.

Analysis

Training dynamics across reinforcement learning steps, showing reward and evaluation anchor scores. — **Training dynamics.** Reward and anchor scores over reinforcement learning steps.

Ablation study across all metrics, isolating artifact-grounded credit assignment and long-horizon training. — **Ablations.** Contributions of artifact-grounded credit assignment and long-horizon training.

Human evaluation results from pairwise comparisons between WebCoder and four opponent models. — **Human evaluation.** Pairwise preferences from annotators with frontend development experience.

BibTeX

@article{dou2026webcoder,
  title   = {WebCoder: Scaling General Frontend Development with
             Artifact-Grounded Agentic Reinforcement Learning},
  author  = {Dou, Zijun and Yang, Zhen and Liu, Xijun and
             Zhou, Dongzhuoran and Zeng, Aohan and Tang, Jie},
  journal = {arXiv preprint arXiv:2610.23104},
  year    = {2026}
}