Extending the findings to additional translation controls

Determine whether the observed limitations of forgetting-mitigation methods during machine-translation fine-tuning extend to additional translation controls, specifically glossary constraints and gender-neutral translation.

Background

The study evaluates machine-translation-specific instruction following through formality control, grammatical gender control framed as masculine-versus-feminine selection, and length control. Its results show that methods preserving general capabilities can nevertheless fail to preserve these translation-specific controls.

The authors explicitly leave unresolved whether the same pattern applies to glossary constraints and gender-neutral translation, which were not evaluated in the experiments. Gender-neutral translation is also identified in the limitations as outside the evaluation because the adopted benchmark treats gender as binary and lacks the required evaluation machinery.