Languages are Modalities: Cross-Lingual Alignment via Encoder Injection

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Agarwal, Rajan, Gupta, Aarush
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912679395852288
author Agarwal, Rajan
Gupta, Aarush
author_facet Agarwal, Rajan
Gupta, Aarush
contents Instruction-tuned Large Language Models (LLMs) underperform on low resource, non-Latin scripts due to tokenizer fragmentation and weak cross-lingual coupling. We present LLINK (Latent Language Injection for Non-English Knowledge), a compute efficient language-as-modality method that conditions an instruction-tuned decoder without changing the tokenizer or retraining the decoder. First, we align sentence embeddings from a frozen multilingual encoder to the decoder's latent embedding space at a reserved position via a lightweight contrastive projector. Second, the vector is expanded into K soft slots and trained with minimal adapters so the frozen decoder consumes the signal. LLINK substantially improves bilingual retrieval and achieves 81.3% preference over the base model and 63.6% over direct fine-tuning in LLM-judged Q&A evaluations. We further find that improvements can be attributed to reduced tokenization inflation and a stronger cross lingual alignment, despite the model having residual weaknesses in numeric fidelity. Treating low resource languages as a modality offers a practical path to stronger cross-lingual alignment in lightweight LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2510_27254
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Languages are Modalities: Cross-Lingual Alignment via Encoder Injection
Agarwal, Rajan
Gupta, Aarush
Computation and Language
Artificial Intelligence
Machine Learning
Instruction-tuned Large Language Models (LLMs) underperform on low resource, non-Latin scripts due to tokenizer fragmentation and weak cross-lingual coupling. We present LLINK (Latent Language Injection for Non-English Knowledge), a compute efficient language-as-modality method that conditions an instruction-tuned decoder without changing the tokenizer or retraining the decoder. First, we align sentence embeddings from a frozen multilingual encoder to the decoder's latent embedding space at a reserved position via a lightweight contrastive projector. Second, the vector is expanded into K soft slots and trained with minimal adapters so the frozen decoder consumes the signal. LLINK substantially improves bilingual retrieval and achieves 81.3% preference over the base model and 63.6% over direct fine-tuning in LLM-judged Q&A evaluations. We further find that improvements can be attributed to reduced tokenization inflation and a stronger cross lingual alignment, despite the model having residual weaknesses in numeric fidelity. Treating low resource languages as a modality offers a practical path to stronger cross-lingual alignment in lightweight LLMs.
title Languages are Modalities: Cross-Lingual Alignment via Encoder Injection
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2510.27254