Cross-linguistic generalization of linear truth directions

Determine whether the linear truth direction identified by probes trained on English medical exam corpora behaves similarly in other languages.

Background

The study evaluates English-language exam corpora, so its cross-corpus comparisons are confounded with differences in exam region. The authors explicitly leave unresolved whether the observed behavior of linear truth directions extends beyond English-language settings.

References

All corpora are English exam corpora, so the cross-corpus comparison confounds corpus with exam region, and whether the same picture holds in other languages is an open question.

— Investigating Linear Probe Robustness to Linguistic Register, Medical Specialty, and Corpus Shifts in Medical QA  (2609.01361 - Mishra et al., 1 Sep 2026) in Section 6, Limitations