Two-class pretraining still supports multiclass table predictions

TabICLv2's synthetic classification tasks are usually capped at ten classes. Does that cap describe only the tasks used for training, or does it also limit how many classes the model can predict?

The answer matters to what this particular training setting can support, and the paper tests the question by retraining with binary tasks.

The authors retrain TabICLv2 on synthetic tasks with only two classes, then test whether context labels still let it predict datasets with more classes.

TabArena is a test suite that compares classification performance across 38 binary and multiclass datasets. Its Elo rating summarizes relative performance across datasets.

In zero-shot evaluation, the binary recipe scored 1,590.1 Elo versus 1,568.0 for the official default, with bagging enabled and no gradient updates or post-hoc calibration.

The bootstrap intervals overlap, so the higher score is directional evidence that binary-prior training can retain multiclass prediction in this test.

A separate credit-scoring evaluation also favored the binary checkpoint on most institutions, but both comparisons involve training-recipe differences.

The TabArena score difference is directional because the bootstrap intervals overlap. The binary recipe also used a shorter, two-stage training schedule than the official default, so the comparison does not show that changing the class count alone caused the result.