Causes of Cross-Linguistic Performance Disparities

Determine whether performance disparities across languages in multilingual language models reflect model biases, data biases, or specific linguistic properties.

Background

The paper situates structural transfer within multilingual language modeling, where models must learn from heterogeneous languages with unequal data distributions. Performance differences across languages may result from properties of the model or training data rather than from inherent linguistic differences. Resolving this question would clarify how to interpret cross-lingual evaluation disparities and how to design more effective multilingual training regimes.

References

An open question is whether performance disparities across languages reflect model or data biases, or instead arise from specific linguistic properties .

— Structural priors for data-efficient language learning  (2609.11505 - Veitsman et al., 10 Sep 2026) in Section 2, paragraph “Multilingual modeling”

It remains an unanswered question which properties of the structural source interact with the target language data and how they influence the performance on the evaluation task.

— Structural priors for data-efficient language learning  (2609.11505 - Veitsman et al., 10 Sep 2026) in Section 2, paragraph “Empirical findings”

Whether the decodability–causality separation and the category-specificity of bias features hold in typologically diverse or lower-resource languages remains unexamined.

— Tracing Stereotypes from Representation to Output in Multilingual LLMs  (2609.08322 - Tumurchuluun et al., 8 Sep 2026) in Limitations section

What causes this breakdown remains open: natural language pairs differ in script, grammar, segmentation, token vocabulary, and data distribution at once, so these measurements alone cannot say which difference is responsible.

— Why Pretraining Fails to Share Cross-Lingual Knowledge  (2609.19291 - Gaber et al., 16 Sep 2026) in Section 3, immediately before Section 4

We leave understanding these factors, possibly utilizing our controlled environment, and improving upon them for future work.

— Why Pretraining Fails to Share Cross-Lingual Knowledge  (2609.19291 - Gaber et al., 16 Sep 2026) in Section 6, Conclusions

It remains to be seen whether the gains from L1-prefix-probability scoring hold for other L2 languages and for L1 backgrounds not represented in these benchmarks, especially those that may be underrepresented in the models' pretraining data.

— Inferring L1 Backgrounds from L2 English with LLM Token Probabilities  (2610.23309 - Uluslu et al., 16 Oct 2026) in Section 'Limitations', paragraph 'Dataset representativeness and low-resource scenarios'