Multilingual Alignment for Enterprise LLMs
Aligning LLMs for Multilingual Consistency studies performance differences between English and other languages in enterprise applications, including systems using retrieval-augmented generation.
Authors: Amit Agarwal, Hansa Meghwani, Hitesh Laxmichand Patel, Tao Sheng, Sujith Ravi, Dan Roth. Venue: EMNLP 2025 Industry Track.
Post-training method
The method places semantically equivalent examples in different languages into the same training batch. Fine-tuning with this structure aligns model behavior across languages and improves non-English accuracy in the study’s evaluation while preserving English performance.
This work connects model alignment to practical multilingual tasks such as customer support, content moderation, and information retrieval. It informs my interest in post-training objectives that improve consistency across languages.
Paper and citation
Explore multilingual reasoning model training and more publications.