Traditional document parsing systems usually output Markdown, which works well for linear text but struggles to preserve fine‑grained spatial structure, table geometry, visual hierarchy, and rendering semantics. HTML can fully describe heterogeneous, layout‑sensitive documents while remaining editable, inspectable, and reusable. To study this setting systematically, we introduce AmbiHTMLBench.
In the benchmark, a model receives a real‑world document and must generate executable HTML that retains its content, organization, and visual appearance. The dataset comprises 1,115 diverse pages with structured annotations. Evaluation follows two protocols: the objective protocol aligns ground‑truth and predicted rendered blocks, measuring table fidelity, page‑level text error, and two‑dimensional layout consistency; the reference‑anchored protocol lets LLM judges align with human preferences to assess perceived fidelity.
Experiments across ten leading multimodal systems show the highest objective score of 0.8092, while the subjective winner differs, indicating that automatic decomposition metrics and human perception capture complementary aspects. Fine‑tuning a small‑parameter expert model for the task further improves document‑to‑HTML reconstruction performance.
These findings position AmbiHTMLBench as a diagnostic testbed for building smarter, more logically consistent document parsing systems.
Review: AmbiHTMLBench provides a systematic evaluation framework for HTML‑centric document parsing, balancing objective metrics with subjective perception and offering clear guidance for future model improvements.