Efficient Low-Resource Language Adaptation via Multi-Source Dynamic Logit Fusion

TL;DR

TriMix is a dynamic logit fusion framework that enhances low-resource language adaptation by balancing specialized small models, high-resource instruction tuning, and large model scaling, without requiring task annotations.

Contribution

Proposes TriMix, a novel test-time logit fusion method that improves low-resource language adaptation by dynamically combining multiple model sources efficiently.

Findings

TriMix outperforms single-model baselines and Proxy Tuning across multiple languages.

Prioritizing small LRL-specialized model logits is key for success.

TriMix requires no task annotations and only continual pretraining on small models.

Abstract

Adapting large language models (LLMs) to low-resource languages (LRLs) is constrained by the scarcity of task data and computational resources. Although Proxy Tuning offers a logit-level strategy for introducing scaling effects, it often fails in LRL settings because the large model's weak LRL competence might overwhelm the knowledge of specialized smaller models. We thus propose TriMix, a test-time logit fusion framework that dynamically balances capabilities from three different sources: LRL competence from a continually pretrained small model, task competence from high-resource language instruction tuning, and the scaling benefits of large models. It is data- and compute-efficient, requiring no LRL task annotations, and only continual pretraining on a small model. Experiments across four model families and eight LRLs show that TriMix consistently outperforms single-model baselines and Proxy Tuning. Our analysis reveals that prioritizing the small LRL-specialized model's logits is crucial for success, challenging the prevalent large-model-dominant assumption.