Paper Overview
Field: NLP Authors: Mengyu Bu, Yang Feng Published: 2025-03-18 arXiv: 2503.13831
Summary
Large language models (LLMs) exhibit strong general intelligence, yet their multilingual performance remains highly imbalanced. Although LLMs encode substantial cross-lingual knowledge in a unified semantic space, they often struggle to reliably interface this knowledge with low-resource or unseen languages. Fortunately, pretrained encoder-decoder translation models already possess balanced multilingual capability, suggesting a natural complement to LLMs.
Approach
The authors propose XBridge, a compositional encoder-LLM-decoder architecture that:
- Offloads multilingual understanding and generation to external pretrained translation models
- Preserves the LLM as an English-centric core for general knowledge processing
- Introduces lightweight cross-model mapping layers to address representation misalignment across models
- Uses an optimal-transport-based alignment objective to achieve fine-grained semantic consistency in multilingual generation
Results
Experiments across four LLMs on multilingual understanding, reasoning, summarization, and generation tasks show that XBridge outperforms strong baselines, particularly on low-resource and previously unseen languages — all without retraining the LLM.
---
*Auto-collected on 2026-03-19*