TABVERSE: Benchmarking Cross-Format Table Understanding in LLMs and VLMs

A new benchmark called TABVERSE isolates how the format of a table — whether HTML, Markdown, LaTeX, or a rendered image — changes the reasoning performance of large language and vision-language models, according to a paper posted to arXiv on June 8, 2026 [1]. The benchmark aligns identical table content across multiple structural formats and rendered images, then evaluates models on three tasks: Question Answering, Structural Understanding Capability, and Structure Reconstruction [1][2]. By holding content constant, the design allows researchers to measure the effect of representation alone, a variable that previous evaluations often conflated with changes in layout or modality [2]. Results indicate that representation choice substantially affects table understanding [1]. Models generally perform better when tables are provided as structured text rather than as rendered images, though the size of the performance gap varies by task, model, and format [2]. HTML emerged as the most robust text format across the tested conditions [1]. Certain tasks remained difficult even for the strongest models. Row-sensitive structural understanding and the reconstruction of syntactically usable LaTeX code posed persistent challenges [2]. These findings suggest that a model’s ability to parse a table’s visual or structural boundaries is not uniform, and that format-specific brittleness can skew benchmark scores if left uncontrolled. The work arrives as the research community increasingly relies on table-reasoning benchmarks to compare LLMs and VLMs, yet lacks a standardized method for separating content difficulty from representation difficulty [2]. The authors argue that table representation should be treated as a first-class evaluation variable, not an incidental detail of dataset construction [1][2]. TABVERSE includes question-category and difficulty tags, which enable fine-grained analysis of where models succeed or fail across formats [2]. The benchmark’s multimodal design also makes it possible to compare text-only LLMs against VLMs on the same underlying data, offering a direct window into how visual grounding helps or hinders table comprehension [1][2]. The paper does not propose a new model architecture or training method. Instead, it supplies a diagnostic instrument for measuring a capability that underpins applications in finance, scientific literature analysis, and data journalism, where tables routinely appear in heterogeneous formats [1][2].