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

Unified evaluation of cross-lingual consistency (CLC) enhancement methods across three model families shows post-training distribution alignment outperforms inference-time interventions. While post-training provides more reliable results, cross-domain transfer remains constrained by output format similarity between source and target tasks.