💬 NLP

Language on Demand, Knowledge at Core: Composing LLMs with Encoder-Decoder Translation Models for Extensible Multilinguality

This paper introduces XBridge, a compositional architecture that enhances the multilingual capabilities of English-centric large language models by offloading language understanding and generation to pretrained encoder-decoder translation models, connected via lightweight mapping layers and an optimal transport-based alignment objective to achieve superior performance on low-resource and unseen languages without retraining the core LLM.

Original authors: Mengyu Bu, Yang Feng

Published 2026-03-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mengyu Bu, Yang Feng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, world-class Chef (the Large Language Model, or LLM) who is an expert at cooking complex, gourmet meals. This Chef knows everything about food, can create new recipes, and solve tricky culinary puzzles. However, there's a catch: this Chef only speaks and understands English fluently. If you ask them to cook a dish using ingredients described in Swahili or Bengali, they get confused. They might understand the idea of the dish, but they can't read the recipe card, nor can they explain the final dish back to you in your local language.

Now, imagine you also have a Super-Translator (the Encoder-Decoder Translation Model). This person speaks 200 languages perfectly. They can read a recipe in Swahili, understand it instantly, and write it back out in Swahili. But, they aren't great chefs. If you ask them to solve a complex math problem or write a creative story, they might struggle because their brain is tuned for translation, not deep reasoning.

The Problem:
Usually, if you want a multilingual AI, you try to teach the Chef to speak 200 languages. This is like forcing the Chef to memorize every language in the world. It's expensive, takes forever, and often makes the Chef forget how to cook their best English dishes. Alternatively, you could just use the Super-Translator to translate everything to English, let the Chef cook, and then translate the result back. But this is clunky, slow, and often loses the "flavor" of the original thought.

The Solution: XBridge
The paper introduces XBridge, which is like building a high-speed, magical bridge between the Chef and the Super-Translator.

Here is how XBridge works, step-by-step:

1. The Setup: A Three-Part Team

Instead of trying to make the Chef speak every language, XBridge keeps the Chef exactly as they are (frozen in English) and adds two new roles:

The Interpreter (Encoder): When you ask a question in Swahili, the Interpreter translates it into "Chef-speak" (English concepts) so the Chef can understand it.
The Chef (LLM): The Chef does the actual thinking, reasoning, and problem-solving in their native English. They don't need to learn new languages; they just need to understand the meaning.
The Presenter (Decoder): Once the Chef has the answer in English, the Presenter takes that answer and translates it back into Swahili (or any other language you want) for you to read.

2. The Secret Sauce: "Optimal Transport"

Here is the tricky part. The Interpreter and the Presenter speak a slightly different "dialect" than the Chef. If you just connect them, the Chef might think "Apple" means "Red Fruit," but the Presenter might think "Apple" means "Tech Company." They get out of sync.

The authors invented a special training method called Optimal Transport. Think of this like a dance instructor for the team.

The dance instructor watches the Chef's moves (English thoughts).
They watch the Presenter's moves (Swahili output).
They gently nudge the Presenter to match the Chef's rhythm perfectly, ensuring that every specific thought the Chef has is represented by the exact right word in the target language. It's not just a rough translation; it's a precise, step-by-step alignment of meaning.

3. The Training Process (The Three Stages)

You don't just plug them in and hope for the best. You train them in three phases, like a sports team practicing:

Phase 1 (The Bridge): Teach the Interpreter and Presenter how to talk to the Chef. They learn to pass messages back and forth without losing meaning.
Phase 2 (The Chef's Input): Teach the Interpreter how to take complex questions (like math problems) in different languages and explain them clearly to the Chef.
Phase 3 (The Presenter's Output): Teach the Presenter how to take the Chef's complex English answers and turn them into beautiful, natural-sounding sentences in the target language.

Why is this a Big Deal?

No Re-Training the Chef: You don't have to spend millions of dollars re-teaching the Chef to speak 200 languages. You just build the bridge.
Low-Resource Languages: This works amazingly well for rare languages (like Swahili or Bengali) where the Chef usually fails. The Super-Translator handles the rare language, and the Chef handles the smart thinking.
Language-on-Demand: You can ask the same question in English, but tell the Presenter to reply in Japanese, then in German, then in Hindi, all without changing the Chef's brain.

The Result

In tests, XBridge performed better than trying to force the Chef to learn everything, and it was much smarter than just using a simple translator. It successfully combined the brainpower of a top-tier English AI with the linguistic flexibility of a professional translator, creating a system that is smart, multilingual, and efficient.

In short: XBridge doesn't make the genius speak every language; it gives the genius a perfect team of translators so they can share their genius with the whole world.

1. Problem Statement

Large Language Models (LLMs) exhibit strong general intelligence and reasoning capabilities but suffer from highly imbalanced multilingual performance. While they perform well in English and high-resource languages, they struggle to reliably understand or generate text in low-resource or unseen languages.

The Limitation: The core issue is not a lack of cross-lingual knowledge within the LLM, but the difficulty of interfacing this knowledge with diverse linguistic representation spaces.
The Gap: Existing solutions often rely on data-level augmentation (translating training data) or encoder-only augmentation (injecting encoder representations into the LLM). These approaches either require costly retraining or fail to support multilingual generation, leaving the output generation English-centric.
The Challenge: Simply inserting a frozen LLM between a pretrained multilingual encoder and decoder (NMT) creates a semantic misalignment. The LLM's internal representation space differs from the NMT model's encoder/decoder spaces, and the LLM's output tokenization often conflicts with the decoder's expectations, leading to generation failures.

2. Methodology: XBridge

The authors propose XBridge, a compositional framework that integrates a frozen, English-centric LLM with a pretrained multilingual Encoder-Decoder Neural Machine Translation (NMT) model.

Architecture

XBridge adopts an Encoder-LLM-Decoder architecture:

Multilingual Encoder: A pretrained NMT encoder (e.g., NLLB) processes input text in any language $L_x$ into a shared semantic representation space.
Mapping Layer (Encoder-side): A lightweight MLP projects the encoder's representations into the LLM's input space.
Frozen LLM: Acts as the English-centric core for general knowledge processing and reasoning. It receives the mapped representations and an English instruction prompt, generating an English output sequence $z$ .
Mapping Layer (Decoder-side): A lightweight MLP projects the LLM's hidden states (specifically from the penultimate layer to preserve semantic richness) into the decoder's representation space.
Multilingual Decoder: A pretrained NMT decoder takes the projected LLM states as keys/values for cross-attention and generates the final output in the target language $L_y$ .

Key Technical Innovations

Optimal Transport (OT) Alignment: To solve the semantic misalignment between the LLM's output space and the decoder's input space (caused by heterogeneous tokenizers and representation spaces), XBridge introduces an OT-based alignment objective.
- It treats the alignment as a soft, many-to-many matching problem between the LLM's English token representations and the re-encoded representations of the same text via the multilingual encoder.
- This ensures token-level semantic consistency without requiring exact positional correspondence, making it robust to sequence length mismatches.
Three-Stage Training Strategy: To ensure stable optimization without retraining the LLM, XBridge uses a decoupled training process:
1. Stage 1 (Cross-Model Mapping): Aligns the encoder, LLM, and decoder using trilingual data ( $x \to z \to y$ ). Only mapping layers and decoder attention are trained.
2. Stage 2 (Encoder-Side Adaptation): Fine-tunes the encoder-side mapping on task-specific instruction data to teach the LLM how to utilize multilingual inputs.
3. Stage 3 (Decoder-Side Adaptation): Fine-tunes the decoder-side mapping and attention layers to optimize multilingual generation quality, leveraging the stable LLM output distribution established in Stage 2.

3. Key Contributions

Compositional Architecture: A novel framework that offloads multilingual understanding and generation to external NMT models while preserving the LLM as a frozen, high-quality reasoning core.
Optimal Transport Alignment: A mechanism to bridge heterogeneous representation spaces between LLMs and NMT models, enabling fine-grained semantic consistency for multilingual generation.
Parameter Efficiency: The method requires training only lightweight mapping layers and the decoder, avoiding the massive computational cost of fine-tuning the entire LLM or retraining on multilingual corpora.
Language-on-Demand Generation: The system supports flexible generation in arbitrary languages by simply switching the target language token in the decoder, without retraining.

4. Experimental Results

The authors evaluated XBridge on four base LLMs (MetaMath-7B, LLaMA3-8B, Aya-23-8B, Qwen2.5-7B) across four tasks: multilingual translation, reasoning (MGSM), summarization (XL-Sum), and commonsense reasoning.

Translation (FLORES-101): XBridge significantly outperforms strong baselines (SFT, Translate-Test, MindMerger, LayAlign), particularly on low-resource languages (e.g., Bengali, Swahili). It approaches the performance of the external NMT model (NLLB-200) while maintaining the LLM's reasoning capabilities.
Reasoning (MGSM): XBridge achieves accuracy comparable to English-only reasoning when the output is generated via the decoder, proving that the reasoning core is preserved and effectively transferred across languages.
Summarization (XL-Sum): XBridge outperforms encoder-augmented baselines and translation-cascaded systems, demonstrating superior faithfulness and fluency in multilingual summaries.
Generalization: The model generalizes well to 42 unseen languages not used during training, confirming that the learned mappings are language-agnostic rather than overfitted to specific training languages.
Ablation Studies: Removing the decoder (encoder-only) fails to support multilingual generation. Removing the OT alignment causes significant degradation in generation quality, especially for languages with different tokenization schemes.

5. Significance

Solving Multilingual Imbalance: XBridge offers a practical solution to the "multilingual curse" of LLMs, enabling high-performance reasoning and generation in low-resource languages without the prohibitive cost of retraining large models.
Modularity and Extensibility: By decoupling the "knowledge core" (LLM) from the "language interface" (NMT), the framework allows for easy updates. One can swap the NMT model for a better one or upgrade the LLM without retraining the entire system.
Efficiency: It achieves near-state-of-the-art multilingual performance with minimal additional parameters and training data, making it highly scalable for real-world deployment in diverse linguistic environments.
Future Direction: This work suggests a paradigm shift from "training larger multilingual models" to "composing specialized models" to achieve extensible and balanced multilingual AI.