JIVEAdapter Decomposes Multi-Task Low-Rank Adapters into Shared and Task-Specific Weights
October 7, 2026
JIVEAdapter introduces an additive low-rank adapter that decomposes weight updates into a shared joint structure and orthogonal per-task individual structures. This architecture allows for adaptive rank allocation and more interpretable separation of shared and task-specific signals during fine-tuning.
HOW THIS AFFECTS YOU
●
builderYou can use this to train more efficient multi-task adapters that reuse shared knowledge across different tasks.
●
researcherThis provides a method to explicitly separate and interpret joint versus individual variation in PEFT.