Can a model build a runnable, interactive front-end artifact that survives end-to-end verification — not just generate pretty HTML?模型能否构建一个可运行、可交互、能通过端到端验证的前端产物，而非只是生成好看的 HTML？
Leaderboard排行榜
How models rank on this subset. Scores are averaged over three runs on the current 70-task Web set.各模型在该子集上的排名。分数为当前 70 道 Web 任务三轮运行的平均结果。
Across both harnesses, the main separator is front-end engineering closure rather than visible page generation. Document conversion and constrained visual/page tasks are stronger; code testing, analysis reports, data semantics and multi-step state closure expose more failures. The common failure modes are stale runtime state or payloads, weak source-to-output evidence for charts and reports, and tests that do not fully bind to real product behavior.两个运行环境下，主要区分点不是能否生成可见页面，而是前端工程闭环能力。文档转换和约束明确的视觉/页面任务相对更稳；前端代码测试、前端分析报告、数据语义和多步骤状态闭环更容易暴露问题。常见失败包括运行时状态或 payload 不闭合、图表和报告缺少 source-to-output 证据链，以及测试没有充分绑定真实产品行为。
Most failed items come from the broad LLM/VLM layer; rule failures usually point to hard delivery-contract misses, and agent failures correspond to broken workflow or state behavior.失败项主要来自覆盖面最广的 LLM/VLM 层；rule 失败通常对应硬性交付契约缺失，agent 失败则对应流程或状态行为没有真正走通。
Model settings: reasoning effort defaults to high, and each model runs with its provider-default inference hyperparameters.模型设置：reasoning effort 默认为 high，各模型使用其默认推理超参。 Harness settings: the context window is unified at 200k with a unified auto-compaction threshold; the WebSearch and AskUserQuestion tools are disabled.harness 设置：上下文窗口统一为 200k，自动压缩阈值统一；禁用 WebSearch 与 AskUserQuestion 工具。 Evaluation is pinned to specific harness builds — codebuddy-code:2.109.3 and claude-code:2.1.187; metrics may shift as harness versions evolve.评测固定在特定 harness 构建上：codebuddy-code:2.109.3 与 claude-code:2.1.187；随 harness 版本演进，指标可能变化。 The HY (Hunyuan) endpoint in this evaluation is served first-party by its provider; the other models are accessed through third-party serving endpoints, whose parameter configuration and request handling may affect metrics.本次评测中 HY（混元）端点由其提供方第一方部署；其余模型通过第三方推理端点接入，第三方的参数配置与请求处理可能影响指标。
Artifact comparison产物对比
Artifacts generated by different models for the same task. The examples cover an interactive page, a 3D rendered page, an advertising design page, and a Web game. Use the tabs at the bottom to switch between model outputs.同一题目在不同模型生成的产物。演示题目覆盖：可交互页面、3D 渲染页面、广告设计页面、网页游戏。点击底部切换查看不同模型的产物结果。
Example 1 / 4
Build a 3D snow-mountain cable-car hut scene with drag rotation and wheel zoom生成一个 3D 的雪山缆车小屋场景，需要支持鼠标拖拽旋转视角和滚轮缩放
Evaluation process评测过程
A representative run shows how the evaluator follows the artifact like a user and turns workflow evidence into rubric item verdicts.下面用一个代表性运行展示评测器如何像用户一样操作产物，并把交互证据转成 rubric item 判定结果。
Dataset overview数据集概览
Web is a benchmark for front-end artifact generation, modification, analysis and quality assurance — 70 tasks covering page implementation, interactive apps, visual design, data visualization, analysis reports, state repair and test generation.
It targets the capability mix behind front-end work: requirement understanding, context handoff, state modelling, data and business semantics, visual quality, test coverage and interactive acceptance. The goal is not just to produce a visible page, but to deliver a runnable artifact that can be checked against user-facing behavior.
Web 是一套面向前端产物生成、修改、分析与质量保障的数据集，共 70 道任务，覆盖页面实现、交互应用、视觉设计、数据可视化、分析报告、状态修复与测试生成。
Existing Web and front-end benchmarks are often strong on one slice: static-page reproduction, interactive generation, browser-agent operation, front-end testing, or real code maintenance. This subset puts those slices in one task space. Its task categories, lifecycle coverage, and interaction complexity are sampled from open-source datasets and real user-request datasets, then designed to approximate the request mix seen in real Web scenarios.已有 Web 与前端基准往往强在某一个截面：静态页面复刻、交互网页生成、浏览器 agent 操作、前端测试或真实代码维护。该子集把这些能力放进同一套任务空间。任务分类、覆盖生命周期和交互难度，都基于开源数据集和真实用户请求数据集采样，模拟真实用户在 Web 场景的真实请求比例进行设计。
1Beyond page generation — the full Web work-surface覆盖完整 Web 工作面，而非仅页面生成
Many Web benchmarks concentrate on single-page generation or visual reproduction. The task face here is wider: 70 tasks split into 7 categories, led by page interaction and data visualization, while keeping front-end project analysis, code testing and document conversion — work that appears around real Web delivery, not just page authoring. Over half the suite lands on interaction and visualization; 19 tasks cover front-end analysis, testing and conversion.很多 Web 基准集中在单页生成或视觉复刻。该子集的任务面更宽：70 道任务切成 7 类主分类，以页面交互和数据可视化为主，同时保留前端项目分析、代码测试与文档转换等真实 Web 交付周边工作，而非仅评测页面编写。超过一半任务落在交互与可视化上；另有 19 道 前端分析、测试与转换类任务。
2The whole dev lifecycle, not just From Scratch覆盖开发生命周期，而非仅从零开发
The subset is not only “can you write a page from zero”. From Scratch is the largest mode but only half the suite (35/70); the other half spreads across bug fix, feature extension, review & analysis, test generation and format conversion — existing-context work that is common in real front-end development.该子集并非单纯测能不能从零写出一个页面。从零开发是最大类，但只占 一半（35/70）；另一半分布在缺陷修复、功能扩展、审查分析、测试生成与格式转换，这些都是实际前端开发中常见的已有上下文工作。
3Interaction and state are first-class requirements交互与状态是一等能力项
Many failures only appear after interaction: filters drift, submitted state is lost, views disagree, or a workflow cannot recover after refresh. The subset explicitly tags this axis: 45 of 70 tasks require some interaction or state behavior, including single-flow state, persistence/offline/cross-state behavior, multi-step workflows and lighter UI state.很多失败只有交互后才会出现：筛选状态漂移、提交状态丢失、不同视图不一致，或刷新后流程无法恢复。该子集显式标注这条轴线：70 道任务中的 45 道包含交互或状态要求，覆盖单流程状态、持久化 / 离线 / 跨状态、多步骤工作流和轻量 UI 状态。
Scoring评测方式
Scoring is rubric-item based. Each task defines scored items, each item maps to one judge type, and each item returns pass or fail before penalties are aggregated into the task score.
Rule judge checks deterministic requirements such as files, output paths, structured fields, test/build signals and task-specific prechecks.
LLM/VLM judge reads text, code, reports, structured evidence and screenshots to judge semantics, layout, visual consistency and information visibility.
Agent judge opens the running artifact, executes the required workflow and checks state changes from a user-like path.
Task score is subtractive: failed items subtract their penalties from 1.0, using the regular 0.1 / 0.2 / 0.3 levels; a fatal failure sets the task score to 0. The final Web score is the mean over the 70 task scores.
“Based on the player feedback, difficulty parameters, and input rules in the attachments, please build a Snake mini-game. The arrow-key controls, difficulty tiers, turn buffering, and speed pacing all need to match the spec — no accidental reverse moves and no dropped inputs.”「根据附件中的玩家反馈、难度参数和输入规则，构建一个贪吃蛇小游戏。方向键控制、难度档位、转向缓冲和速度节奏都需要符合规格，不能出现误反向移动或输入丢失。」
Rubric itemsRubric 列表
# snake-gameplay-implementation-L3-010
llm difficulty_modes_present four selectable difficulty tiers
llm timing_bands_match movement timing follows the parameter table
agent input_buffer_rules legal double turns execute; reverse moves are rejected
llm deterministic_trace_stable fresh rounds follow the supplied baseline trace
agent self_collision_excludes_tail collision logic allows moving into the old tail cell
agent best_score_persists_across_reload best score is displayed and restored after reload
Related work相关工作
Compared with representative Web and front-end benchmarks: visual/page reproduction, interactive page generation, browser-agent operation, and coding-agent workflow benchmarks. Our benchmark covers a broader combination of task types, development modes, interaction/state behavior, and rule, LLM/VLM, and agent judges.与代表性 Web 和前端基准对比，包括视觉/页面复刻、交互页面生成、浏览器 agent 操作和 coding-agent 工作流基准。我们的基准在任务类型、开发模式、交互/状态覆盖，以及 rule、LLM/VLM、agent judge 组合上覆盖更全面。