TableVista benchmark finds foundation models falter on complex tables rendered as images
A new benchmark of 30,000 multimodal table-reasoning samples shows 29 leading models stay stable across rendering styles but degrade sharply on complex layouts and vision-only inputs.
In plain words
Researchers built TableVista, a test set for checking how well AI models answer questions about tables presented as images rather than plain text. Each of its 3,000 problems was turned into ten visual versions, creating 30,000 samples. Testing 29 models showed they handle style changes well but struggle when table layouts are complex or the table exists only as a picture.
What was announced
The benchmark's creators assembled 3,000 table-reasoning problems and ran each through a rendering and transformation pipeline that produced ten visual variants per instance, spanning different scenario styles, robustness perturbations, and vision-only configurations. This expansion yielded 30,000 multimodal samples intended for multi-dimensional evaluation of foundation models, the term for large general-purpose AI systems that can process mixed inputs such as text and images.
In evaluations covering 29 state-of-the-art open-source and proprietary models, the authors report that performance remained largely stable across diverse rendering styles. However, they observed pronounced degradation on complex structural layouts and in vision-only settings, concluding that current models struggle to keep reasoning consistent when structural complexity combines with visually integrated presentations. The team presents this as exposing critical gaps in multimodal capability and as guidance for building more reliable table-understanding models. The findings come from the benchmark's own evaluation suite, so they measure behavior on TableVista's tasks rather than independent confirmation of model weaknesses in other settings.
What it does not show
- Evidence is limited to the paper's abstract, which describes the benchmark and headline findings but not the full methodology, per-model results, or error analysis.
- The reported degradation reflects performance on TableVista's own tasks and 29 evaluated models; the abstract does not establish how these results generalize to other benchmarks or real-world deployments.
- No independent party has replicated the evaluation; the findings are the authors' own analysis.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.