Resolve TabPFN v3’s directional response asymmetry

Determine whether TabPFN v3 exhibits unequal responses to negative-to-positive versus positive-to-negative dependence reversals in the evaluated return-forecasting experiments.

Background

The paper compares the history penalty produced by opposite directions of dependence change while keeping the recent observations and conditional distribution of the next return matched. Across the directional contrasts, 16 of 20 positive contrasts are established within each process, but TabPFN v3’s four directional contrasts remain unresolved.

The unresolved contrasts do not establish that TabPFN v3 is invariant to reversal direction, and the paper emphasizes that they also do not identify the source of any potential asymmetry in pretraining, normalization, representation, or the output rule.

References

Direct paired comparisons establish 16 positive directional contrasts out of 20 within each process, favoring negative-to-positive reversals. TabPFN's four directional contrasts remain unresolved. This is evidence of unequal responses for the resolved comparisons, rather than evidence that the unresolved model is invariant.

— Conflicting History in Foundation Model Return Forecasts  (2610.25822 - Bulut, 19 Oct 2026) in Section 3.2, page 6