Research map / Vision-language and multimodal / Multimodal LLMs

ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes creative table visualization and the ShowTable pipeline, where an MLLM rewrites tables into visual plans and reflects on generated infographics while a diffusion editor refines them via RL-trained refinement, evaluated on the new TableVisBench.

Problem

Generating infographics that are both aesthetic and faithful to the data in a given table is hard for current image generation and unified models, which make data-mapping errors in bar heights, labels, and proportions.

Approach

A progressive self-correcting loop in which an MLLM plans and rewrites the table into a visual description, a diffusion model generates an initial image, the MLLM audits it against the table and issues edit instructions, and a diffusion editor refines it; the refinement editor is trained with GRPO using a pairwise-trained reward model.

Key results

• Full ShowTable pipeline with Blip3o-Next raises the TableVisBench score from 10.8 to 34.8 and Bagel from 10.1 to 32.7.
• With Qwen-Image as generator, RW+REF reaches 54.9 overall versus 44.3 for the base model.
• Using Wan2.5-I2I-Preview as the refiner reaches 63.4 overall on TableVisBench, versus 62.5 for Wan2.5-T2I alone.

Surprising findings

• Refinement with the base Qwen-Image-Edit-2509 degrades performance with each correction round (54.3 to 49.4), while the RL-trained editor improves across rounds.
• Code-based chart generation with Gemini-2.5-pro achieves higher data accuracy and relative relationship scores than the ShowTable pipeline (score 79.4 vs 67.9).

Limitations

• Pipeline relies on closed proprietary MLLMs (GPT-5, Gemini-2.5-pro) for reflection, annotation and data construction.
• Base editing models can degrade over iterative correction rounds, so performance depends on the refiner's capability.
• Code-based baselines remain more numerically accurate than the generative pipeline.

Experimental setup

Datasets: SlideVQA OpenImages Cambrian-10M

Benchmarks: TableVisBench

Metrics: Data Accuracy Text Rendering Relative Relationship Additional information Accuracy Aesthetic Quality composite score

Baselines: Flux Bagel Blip3o-Next UniWorld-V1 OmniGen2 Qwen-Image Code (Gemini-2.5-pro matplotlib)

Builds on: GRPO Flow-GRPO Bradley-Terry reward modeling ImageReward

Closest papers by meaning

• TABLET: A Large-Scale Dataset for Robust Visual Table Understanding (ICLR 2026 · similarity 0.88)
• Factuality Matters: When Image Generation and Editing Meet Structured Visuals (ICLR 2026 · similarity 0.86)
• SpatialViz-Bench: A Cognitively-Grounded Benchmark for Diagnosing Spatial Visualization in MLLMs (ICLR 2026 · similarity 0.86)
• VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations (ICLR 2026 · similarity 0.85)
• TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective (CVPR 2026 · similarity 0.85)
• TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.84)