Multilingual language models often produce inconsistent answers to semantically equivalent questions across languages. This work evaluates methods to enhance cross-lingual consistency (CLC) across three model families and three benchmarks.
The study finds that post-training methods, especially direct distribution alignment, reliably improve CLC across different model-dataset combinations. Other methods tend to be more sensitive to answer format and language coverage, and cross-domain transfer remains limited unless source and target tasks share similar output formats.
The evaluation also considers whether CLC enhancement affects models' ability to respond differently when culturally appropriate. Results indicate no systematic degradation in controlled, closed-form tasks, but open-ended responses, especially in non-English, can show occasional accuracy reductions.
This work informs future development of post-training techniques and benchmarks, emphasizing the importance of evaluating both cross-domain robustness and cultural appropriateness.
Source: https://arxiv.org/abs/2609.04409