Xiaomi trains a tabular model on fake data only

Xiaomi built Xiaomi-TabLDM, a tabular model trained only on synthetic causal data, with no real spreadsheets in its training set. Tabular data, pricing tables, sensor logs, transaction records, usually can't leave the company that owns it. A model that never needs real examples sidesteps that privacy problem entirely.

It ranks first on OpenML-CTR23 and second on regression across three separate benchmark suites: TALENT, TabArena and BCCO. On TabArena regression, it posts the second-highest Elo score behind TabFM. Getting there costs 82% less training time and 68% less prediction time than TabFM needs.

The paper doesn't say whether this generalizes past benchmark tables shaped by clean causal structure. Structural causal models are a specific bet about how spreadsheet columns relate, and a second-place finish on TabArena regression means Xiaomi hasn't beaten TabFM outright. Whether the same synthetic-only approach holds up on tables with messier, non-causal relationships, the kind production data usually has, is still open.