Intel / article

A Systematic Evaluation of Cross-Lingual Consistency Enhancement Methods in Multilingual Language Models

Source: arxiv.org
Author: Jirui Qi, Mingyang Wang, Hinrich Sch\"utze, Raquel Fern\'andez, Arianna Bisazza
Date: Published Sep 7, 2026

Context

Multilingual models often answer the same question differently depending on the language, and methods that make answers consistent across languages are hard to compare because they are usually tested under different setups. A comparative study tested several such methods on question answering across three model families and found training-based approaches generally more reliable than interventions applied while answering. Improvements transferred poorly between tasks with different answer formats. The authors also checked whether consistency training hurts answers that should differ by culture. On fixed-answer cultural questions they found no systematic accuracy loss, while open-ended responses sometimes became less accurate, particularly outside English. The findings make answer format part of the : consistency measured on fixed-response tests does not establish culturally appropriate answers in conversation.

Terms in this piece · Glossary

eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source arxiv.org