JevSoup: Training-Free LoRA Composition via System-One Routing
September 28, 2026
JevSoup is a training-free framework that routes between LoRA experts using only input and expert descriptions. It combines experts by projecting the second expert onto the orthogonal complement of the first expert's row space, yielding up to 1.21% accuracy gains on Qwen3 scales.
HOW THIS AFFECTS YOU
●
builderYou can compose multiple specialized LoRA experts at inference time without additional training or complex routers.
●
researcherThis approach explores orthogonal projections as a way to merge modular weights without interference.