New Framework Improves LLM Understanding of Complex Tables with Hierarchical Structure

Large language models struggle with complex tables that contain multi-level headers, merged cells, and irregular layouts. Existing approaches—like table linearization or grid normalization—often fail to capture the hierarchical relationships that give tables their meaning.

A new research paper from researchers proposes the Orthogonal Hierarchical Decomposition (OHD) framework to address this challenge. Rather than flattening tables into sequences or forcing them into normalized grids, OHD preserves structural information.

How It Works#

The framework introduces the Orthogonal Tree Induction (OTI) method, which decomposes irregular tables into two separate trees: one capturing vertical (column) hierarchies and another capturing horizontal (row) hierarchies. This dual-tree approach respects the spatial and semantic relationships within the table itself.

Once the table structure is decomposed, the system uses a "dual-pathway association protocol" to reconstruct the semantic lineage of each cell—essentially tracing how each data point relates to its row and column headers. An LLM then acts as a "semantic arbitrator" to align information across multiple hierarchy levels.

Testing and Results#

The researchers evaluated OHD on two benchmarks for complex table question-answering: AITQA and HiTab. The framework consistently outperformed existing representation methods across multiple metrics, suggesting that preserving table structure leads to better reasoning and understanding.

This approach could improve how AI systems handle real-world data—from financial reports and spreadsheets to scientific tables with complex layouts.