Research map / Vision-language and multimodal / Multimodal LLMs

TABLET: A Large-Scale Dataset for Robust Visual Table Understanding

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

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.

Introduces TABLET, a 4M-example visual table understanding dataset built from 2M tables (88% with original visualizations), plus the VisualTableQA benchmark, and shows SFT of Qwen2.5-VL-7B on it improves held-in/held-out VTU performance and robustness to real-world table images.

Problem

Visual table understanding lacks large-scale training data built from real-world table visualizations; existing VTU datasets use synthetic renderings of serialized tables, fixed examples, or single tasks, causing train-test mismatch.

Approach

Repurpose 14 seed table datasets into a unified instruction format while retrieving original Wikipedia table visualizations (via archived revisions and edit-distance matching) and providing HTML serializations for traceability and re-rendering, plus a human-annotated visual table QA benchmark.

Key results

• Qwen2.5-VL-7B fine-tuned on TABLET-L outperforms zero-/one-shot Qwen2.5-VL-7B on 12 of 14 held-in and held-out tasks, and reaches 84.5 TabMWP and 30.4 ToTTo BLEU.
• On VisualTableQA, TABLET-M achieves 47.8% accuracy vs 42.4% zero-shot (about 13% relative gain) without VTQA training examples.
• Training on original visualizations reduces zero-shot degradation on original vs synthetic tables from 28.9 to 6.63 DegScore points (TABLET-B org).

Surprising findings

• Zero-shot Qwen2.5-VL-7B loses up to 11 points on TabFact when evaluated on original rather than synthetic table images, suggesting synthetic renders overstate VLM performance.
• Including Table Interpretation tasks (column type, entity linking, relation extraction) improves held-in and held-out VTU performance, transferring benefits from text-only table understanding.

Limitations

• Visual table QA benchmark is small (306 examples) and human-annotated.
• Many held-out and test sets may have been seen during pretraining of the evaluated VLMs.
• Original visualizations were only retrievable for about 61.5% of Wikipedia-based tables; remaining tables are synthetic renders.

Experimental setup

Datasets: TURL ToTTo TabFact WikiTableQuestions HybridQA HiTab PubTabNet TabMWP TAT-QA InfoTabs WikiBIO FeTaQA

Benchmarks: WikiBio ToTTo WikiTableQuestions TabMWP HiTab TabFact FeTaQA TAT-QA InfoTabs TabMCQ AIT-QA PubHealthTab

Metrics: exact match accuracy BLEU F1 tree-edit-distance similarity DegScore

Baselines: Qwen2.5-VL-7B zero-shot MMTab fine-tuned Qwen2.5-VL-7B Table-LLaVA 7B DocOwl2 8B InternVL3 8B InternVL3 14B Gemma-3 12B

Builds on: MMTab DocStruct4M WikiDT

Closest papers by meaning

• TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.92)
• VT-Bench: A Unified Benchmark for Visual-Tabular Multi-Modal Learning (ICML 2026 · similarity 0.88)
• TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.88)
• ShowTable: Unlocking Creative Table Visualization with Collaborative Reflection and Refinement (CVPR 2026 · similarity 0.88)
• Multimodal Tabular Reasoning with Privileged Structured Information (NeurIPS 2025 · similarity 0.87)
• TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective (CVPR 2026 · similarity 0.87)