TableNet A Large-Scale Table Dataset with LLM-Powered Autonomous

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

TableNet introduces a new large-scale dataset for Table Structure Recognition (TSR) and an LLM-powered autonomous multi-agent system for table generation and recognition. This system addresses limitations in existing datasets, which often lack scale, diversity, and annotation quality, hindering the effective use of LLMs for complex table layouts. The generation component integrates controllable visual, structural, and semantic parameters to synthesize a wide array of semantically coherent table images with annotations, supporting large-scale dataset construction. The recognition component employs a diversity-based active learning paradigm, selectively sampling informative data to fine-tune a model. This approach achieves competitive performance on the TableNet test set with significantly fewer training samples and superior performance on real-world web-crawled tables compared to models trained on predominant datasets. TableNet includes agent-generated, web-crawled, and augmented open-source tables, ensuring diversity in visual style, structure, and semantics.

Key takeaway

For research scientists developing Table Structure Recognition models, you should consider integrating TableNet or similar autonomously generated datasets. This approach provides significantly more diverse and structurally complex training data, which is crucial for improving real-world generalization. Leveraging diversity-based active learning, as demonstrated, can also substantially reduce the required training samples while maintaining or improving performance on unseen, complex table structures.

Key insights

An LLM-powered multi-agent system autonomously generates diverse, high-quality table datasets for robust Table Structure Recognition.

Principles

• Autonomous generation enhances dataset scale and diversity.
• Active learning optimizes model training with fewer samples.

Method

The system decomposes table generation into schema planning, layout construction, and content filling, using a core LLM, topic/header/body infilling LLMs, and tool modules for CSS, HTML, validation, and rendering.

In practice

• Use multi-agent systems for controlled data synthesis.
• Apply diversity-based active learning for efficient model training.

Topics

• Table Structure Recognition
• LLM-Powered Multi-Agent Systems
• Autonomous Data Generation
• Diversity-Based Active Learning
• Large-Scale Table Datasets

Code references

• WenmuZhou/TableGeneration

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• TRivia: Teach a 3B Model to Parse Tables — Labeled data scarcity and privacy concerns hinder open-source table recognition model performance against proprietary systems.
• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding — Code-guided synthetic data generation significantly improves VLM chart understanding and reasoning capabilities.
• ARC Prize 2025 Paper Award 1st Place TRM — TRM demonstrates that small, recursive models can achieve strong generalization on tiny datasets by learning to iteratively improve answers.
• Synthetic Contrastive Reasoning for Multi-Table Q&A — Synthetic contrastive reasoning traces significantly improve multi-table Q&A performance in LLMs.
• Winning a Kaggle Competition with Generative AI–Assisted Coding — LLM agents, paired with GPU acceleration, dramatically speed up ML experimentation by automating code generation and execution.
• Revisiting Structural Dependency in Autoregressive Multi-Task Table Recognition via Order-Independent Cell-Level Representations — Order-independent cell representations, achieved via non-causal attention, enhance multi-task table recognition and reduce inference time.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.