Impact of Source-Only Fine-Tuning on Target Performance

Establish whether fine-tuning large language models solely on source-language splits of Year-ECLeKTic or similar knowledge-intensive datasets increases target-language accuracy and source–target agreement without any training on target-language data.

Background

Based on their theoretical implications, the authors hypothesize that improving source confidence via overfitting could reduce cross-lingual gaps; they perform preliminary fine-tuning on the source split but do not achieve conclusive evidence.

This leaves open the question of whether source-only fine-tuning reliably yields improvements in target accuracy and agreement, motivating more controlled and powerful experiments.

References

As a result, we do not have a conclusive evidence of the question we attempted to validate: improving source alone also improves target?

— Rethinking Cross-lingual Gaps from a Statistical Viewpoint  (2510.15551 - Piratla et al., 17 Oct 2025) in Appendix: Fine-Tuning Experiments (Section: appendix:sft_expts)

These results are therefore not strong enough to determine whether fine-tuning causes cross-lingual transfer of linguistic abilities.

— The Interlingua Hypothesis: LLMs Translate via a Latent Task-agnostic Feature Space  (2609.00515 - Brinton et al., 1 Sep 2026) in Appendix, Section Additional Fine-Tuning Results, subsection Preservation of monolingual abilities