Large Language Models (LLMs) are fundamentally designed to process one-dimensional sequences of text. However, much of the world’s high-value information—financial reports, scientific data, and web databases—resides in two-dimensional tables. When we feed a table into a model, we typically "linearize" it, flattening the grid into a string of Markdown, CSV, or HTML.
This flattening process creates a phenomenon often described as structural amnesia. By forcing a 2D grid into a 1D line, we destroy the spatial relationships that give the data meaning. Researchers are currently developing new architectures and training paradigms to bridge this "structural gap," ensuring models can "see" the grid even when they are reading a sequence.
The Cost of Linearization
In a standard sequence-to-sequence model, a table is just a long list of tokens. As Tianyu Liu and colleagues noted in Table-to-text Generation by Structure-aware Seq2seq Learning, this approach causes the model to lose the vital connection between a "Field" (the header) and its "Value" (the cell).
The intuition is simple: in a table, a cell’s meaning is defined by its neighbors. A value like "$500" is meaningless without the horizontal context of the "Price" header and the vertical context of the "Product" row. Hiroshi Iida and Mohit Iyyer argue in TABBIE that this flattening destroys the "grid-based logic" required for reasoning. When a table is long, the header might be hundreds of tokens away from the relevant row, exceeding the model's ability to maintain a coherent relationship between them.
Injecting Geometry into Attention
To fix this, researchers have moved toward structure-aware attention. Instead of allowing every token to attend to every other token equally, these models use custom attention masks to "inject" the table's geometry.
In Table Fact Verification with Structure-Aware Transformer, Hongzhi Zhang and his team demonstrated that by explicitly telling the Transformer which cells share a row or column, the model can verify complex claims with much higher accuracy. Similarly, Miao Chen and Tong Xu proposed the TASD framework, which uses hierarchical attention to capture the layout before generating text, effectively teaching the model to "deliberate" on the structure before speaking.
Tables as Graphs and Hypergraphs
If a 1D sequence is too restrictive, perhaps the table shouldn't be a sequence at all. A growing body of work treats tables as graphs, where cells are nodes and row/column relationships are edges.
Xingyao Zhang and Daxin Jiang introduced GraSSLM, which uses graph-based pre-training to capture vertical context that linear strings miss. Taking this further, Sirui Huang and Qing Li proposed HyperG, which uses hypergraphs to model the complex, overlapping dependencies in structured knowledge. By representing a row as a "hyperedge" connecting multiple cells, the model can maintain the integrity of the record even when the text is sparse or noisy.
The Multimodal Recovery
The most recent frontier in this area is hybrid-modal reasoning. Even if text-based serialization is flawed, the visual layout of a table—its lines, colors, and alignments—contains rich semantic cues.
Zhenghao Liu and his team developed HIPPO, a framework that uses Direct Preference Optimization to help Multimodal LLMs extract complementary info from both text and images. This is particularly crucial for "hierarchical" tables—the complex grids found in financial 10-K filings where headers are nested and cells are merged. As Haobo Wang and Junbo Zhao show with their TreeThinker pipeline, understanding these hierarchies requires transforming messy layouts into structured trees, preventing the "hallucinations" that occur when a model loses track of which header belongs to which value.
Why It Works: Intuition
The success of these methods stems from a single realization: structure is a form of compression. A table is a dense map of relationships. When we linearize it, we expand that map into a long, redundant path. By using structure-aware attention, graphs, or visual cues, we allow the model to take "shortcuts" across the data, directly linking a cell to its context without wading through the intervening tokens. This preserves the semantic density of the original data, enabling the model to reason about relationships rather than just sequences.