Table Recognition
NGTR benchmark plus Neighbor-Guided Toolchain Reasoner improve table structure parsing for low-quality inputs.
NGTR · IJCAI 2025 → GitHub →Table Mining Initiative
The USTC AGI Table Mining Group designs benchmarks, toolchains, and autonomous agents that push the boundaries of table understanding, time-series reasoning, and multimodal analytics.
Research Pillars
We connect perception, language, and reasoning to make tables first class citizens for AGI systems.
NGTR benchmark plus Neighbor-Guided Toolchain Reasoner improve table structure parsing for low-quality inputs.
NGTR · IJCAI 2025 → GitHub →TableTime reformulates multivariate time-series classification as table understanding with training-free LLM reasoning.
TableTime · CIKM 2025 → GitHub →STaR develops slow-thinking LLMs with uncertainty-aware inference for stable multi-hop table reasoning.
STaR · Preprint →PoTable and TableMind deliver plan-then-execute reasoning and tool-augmented agents for trustworthy analytics.
PoTable · Preprint → TableMind · WSDM 2026 → TableMind GitHub →ChemTable captures multimodal chemical tables to assess recognition and reasoning across scientific workflows.
ChemTable · Preprint → GitHub →A sweeping review of table mining with LLMs surfaces challenges, advances, and future opportunities.
Table Mining Survey →Flagship Works
2026
Autonomous programmatic agent with tool-augmented reasoning, RAPO training, and sandboxed execution for reliable analytics.
2025
Hierarchical benchmark plus neighbor-guided toolchain to fix low-quality table images and boost VLLM recognition.
2025
Converts multivariate time series to textual tables, enabling zero-shot classification with LLM reasoning pipelines.
2024+
Stage-oriented plan-execute agents and slow-thinking LLMs create transparent, uncertainty-aware reasoning trajectories.
Resources
Structured evaluation pipeline, visual toolchain, and reflection module for robust table recognition.
Prompt templates, neighborhood assistance, and multi-path inference ready for zero-shot classification.
Plan-then-execute operation sets with code generation, sandbox execution, and feedback monitoring.
Expert-curated chemical tables with cell polygons, logical layouts, and domain labels.
Tables are the connective tissue of scientific reports, enterprise systems, and financial decision making. We elevate tables to become native reasoning artifacts for AGI, empowering models with precise structure, semantic context, and executable toolchains.
Multi-stage workflows provide interpretable reasoning plans and verifiable outputs for analysts.
Benchmarks mirror messy real-world tables so agents stay robust beyond curated datasets.
We share toolchains, prompts, and datasets to accelerate community progress in structured reasoning.