Including inverted negative examples in TI

Investigate transitive-inference training that also includes inverted negative examples formed by reversing relations within a branch (for example, s_2^n → s_1^n), and characterize model behavior and generalization when such negatives are incorporated.

Background

The TI setup samples negative examples primarily across different branches. Another natural negative construction reverses within-branch relations (inversions).

The authors omit this case in their experiments and explicitly leave its exploration for future work, indicating an unresolved aspect of task design and its effect on model learning and generalization.

References

Note, another valid way to construct a negative example would be to reverse a relation (e.g. $s_2n \rightarrow s_1n$). We omit this case for simplicity, finding it unnecessary to reproduce the generalization characteristics we observed in PITA, and leave its exploration for future work.

— Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Traces  (2602.14404 - Tong et al., 16 Feb 2026) in Appendix: Task and model details, TI task details