Generalization of interlingua evidence across language models

Investigate whether the observed relationships between monolingual competence, multilingual translation mechanisms, and translation performance hold across a wider variety of language models and whether recent developments in thinking models affect these findings.

Background

The empirical analysis is restricted to Llama-3.1-8B, Aya-23-8B, and TinyAya-3B. Consequently, it remains unresolved whether the reported evidence for the interlingua hypothesis is specific to these model families and sizes or generalizes to other architectures and scales. The authors also identify the possible influence of reasoning-oriented or thinking models as an untested factor.

References

Future work should investigate whether similar trends hold for a wider variety of LLMs, and whether recent developments in thinking models affect these findings.

— The Interlingua Hypothesis: LLMs Translate via a Latent Task-agnostic Feature Space  (2609.00515 - Brinton et al., 1 Sep 2026) in Section Limitations

Its persistence in mathematics-specialised models, under instruction across models and on register-varied pairs is left to future work.

— Objects Without Morphisms: What LLMs for Mathematics Do Not Represent  (2610.03551 - Wang et al., 2 Oct 2026) in Conclusion, final paragraph