Tabular machine learning often comes with a familiar chore: fit a new model for each dataset. Xiaomi-TabLDM is designed to skip that step, using labeled examples as context to predict values for new rows. The trade-off is that avoiding task-specific training does not make inference cheap at every scale.
Using examples as context
Xiaomi released two versions of TabLDM: a 70-million-parameter classifier and a 71-million-parameter regressor. Both are intended to work without separate training for each prediction task.
The model is aimed at tables ranging from hundreds to tens of thousands of rows, with anything from a handful to more than 100 columns. By default, it combines predictions from eight runs. Each run uses a different mix of feature shuffling and data preprocessing, so the final result is not based on a single pass through the table.
The basic idea is to let the labeled data provide the context for a new prediction, rather than first fitting a task-specific model. That makes the approach distinct from a workflow where every new table gets its own training run. It does not, however, remove the need to process the table at prediction time.
How the model processes a table
TabLDM’s architecture has three stages. First, a Set Transformer encodes groups of columns. The model then combines those representations into a vector for each row, before processing rows in the context of the training examples.
Test rows can draw on the training rows, but not on one another. In the final eight blocks, the model uses a mixture-of-experts design. For each token, it selects one of two routed experts along with a shared expert.
A key-value cache is available to speed up repeated requests. Larger tables can be held in memory or stored on disk. These are practical provisions for serving the model, not a promise that table size stops mattering.
Strong benchmark results, with a notable caveat
On the 13 regression datasets in TabArena, Xiaomi-TabLDM scored 1,900 Elo points. That put it behind TabFM, which scored 2,019. Across the full ranking of 51 datasets, TabLDM placed fourth with 1,659 points, close to AutoGluon 1.5 at 1,662 and TabPFN-3 at 1,650.
The developers report that TabLDM’s fit stage runs 82% faster than TabFM’s, while prediction runs 68% faster. Those figures make the model look competitive on both benchmark position and speed, though the reported results do not make it the top performer across the board.
The main limitation is inference cost. It grows quadratically with the number of rows and also with the number of columns, so increasing either dimension can make prediction substantially more expensive. For a model intended to handle tables with tens of thousands of rows, that is not a footnote so much as a condition of use.
An alternative, not a free lunch
Xiaomi-TabLDM is available under the Apache 2.0 license, which allows it to be used and modified. Its context-based approach offers another option for tabular prediction, especially where fitting a separate model for each task is an unwelcome part of the workflow.
The model can spare users a training step. It cannot spare them from thinking about table size, memory, and inference time. In machine learning, the bill rarely disappears. Sometimes it just arrives at a different stage.