MODL: Cross-Lingual Remote-Sensing MLLMs via Orthogonal LoRA Composition
September 23, 2026
The MODL method enables remote-sensing MLLMs to support new languages without paired multilingual RS data. It utilizes a joint loss term to ensure domain and language LoRA updates remain mutually orthogonal, preventing the degradation of the base model's multilingual capabilities.
HOW THIS AFFECTS YOU
●
builderYou can extend English-centric vision-language models to new languages using only text-only multilingual data.
●
researcherThe mutual orthogonality constraint effectively solves the interference problem seen in standard LoRA merging.