A new arXiv paper studies how the text serialization of tables affects the ability of instruction-tuned language models to perform table lookup. The authors design a diagnostic that isolates a common operation: given a table, find the row containing a specified value under one named column, then return the value under another named column.
The results indicate that the way a table is converted into text—its serialization—changes which relations the model can preserve. In other words, the same table presented in different textual formats can lead to different lookup accuracy, even though the underlying data is identical.
The abstract does not specify which serialization formats were tested or which performed best, so the main takeaway is that format is not neutral. For anyone building systems that rely on language models reading tables, the choice of serialization is a meaningful design decision that can affect the model's ability to connect values across columns.