Distilling Multilingual Vision-Language Models: When Smaller Models Stay Multilingual
This addresses the challenge of maintaining multilingual performance in compressed models for vision-language tasks, though it is an incremental study focusing on empirical analysis of distillation approaches.
The paper tackled the problem of multilingual vision-language models performing unevenly across languages when compressed, and found that certain knowledge distillation configurations can preserve or improve multilingual retrieval robustness despite halving model size, while others fail to maintain cross-task stability.
Vision-language models (VLMs) exhibit uneven performance across languages, a problem that is often exacerbated when the model size is reduced. While Knowledge distillation (KD) demonstrates promising results in transferring knowledge from larger to smaller VLMs, applying KD in multilingualism is an underexplored area. This paper presents a controlled empirical study of KD behavior across five distillation approaches, isolating their effects on cross-lingual representation consistency and downstream performance stability under model compression. We study five distillation formulations across CLIP and SigLIP2, and evaluate them on in-domain retrieval and out-of-domain visual QA. We find that some configurations preserve or even improve multilingual retrieval robustness despite halving model size, but others fail to maintain cross-task stability, exposing design-sensitive trade-offs that aggregate accuracy alone does not reveal.