Benchmarking Multimodal LLMs on Code Generation for Complex Interactive Webpages

Researchers introduced WebIGBench, the first benchmark for evaluating multimodal LLMs on code generation for interactive webpages, addressing a critical gap in existing evaluation frameworks that only assess static pages. The benchmark includes 103 real-world webpages with 871 distinct interactive actions and proposes novel automated assessment methods to measure interaction consistency beyond visual fidelity.

WebIGBench represents a significant advancement in evaluating AI-assisted web development tools. Current multimodal LLMs have demonstrated impressive capabilities in converting visual designs to code, but existing benchmarks fail to capture the complexity of modern interactive web applications. This research closes that gap by establishing the first evaluation framework specifically designed for interactive webpage generation, which mirrors real-world development requirements where user interactions and dynamic behaviors are paramount.

The research emerges from a broader trend of AI-driven frontend automation gaining traction in software development. As MLLMs become more sophisticated, the evaluation methods must evolve alongside them to ensure these tools can handle production-level complexity. Traditional metrics focusing solely on visual similarity miss critical functionality—whether generated code properly responds to clicks, form inputs, and other user actions. WebIGBench's collection of 103 complex webpages with 871 distinct interactive actions provides a realistic testing ground that development teams actually need.

For the developer and software engineering communities, this benchmark enables more honest assessment of MLLM capabilities and limitations in practical scenarios. Organizations considering AI-assisted development tools gain clearer visibility into whether these models can truly replace or augment human developers for interactive applications. The research identifies performance boundaries that inform realistic expectations about current model capabilities. The availability of the benchmark as an open resource accelerates the entire industry's ability to improve interactive webpage code generation, driving competition among MLLM providers to enhance their models' interaction handling. This standardization in evaluation metrics also helps establish best practices for assessing AI-generated code quality.

• →WebIGBench is the first benchmark specifically evaluating interactive webpage code generation, covering 871 distinct interactive actions across 5 action types.
• →Existing benchmarks overlook interaction consistency, focusing only on visual fidelity and code structure rather than functional correctness.
• →The research reveals current MLLM performance boundaries in generating executable, interactive code for real-world webpages.
• →Novel automated evaluation pipeline addresses the gap in assessing whether generated code handles user interactions correctly.
• →Open-source benchmark availability accelerates industry-wide improvements in AI-assisted frontend development tools.