Does Serialisation Format Affect LLM Performance on Tabular Classification? A Benchmarking Study

While building an LLM classification pipeline, I noticed that converting tabularrows into text involves a design choice that often gets ignored — which format touse. This made me wonder whether that choice actually affects performance.In this study, I test three serialization formats — natural language sentences,JSON, and Markdown tables — using Flan-T5-XL across three datasets: AdultIncome, Heart Disease, and Bank Marketing. Results show that format choicedoes matter, with differences of up to 14.4 percentage points observed on theHeart Disease dataset. No single format consistently wins, suggesting that se-rialization format deserves more attention as a design parameter in LLM-basedtabular pipelines.