LSA-PAST Reduces Parameter Drift in Compositional Incremental Learning
October 21, 2026
LSA-PAST uses Low-Rank Adaptation (LoRA) and prototype-anchored adapter-space transport to address compositional incremental learning. The method prevents representation drift by stabilizing shared primitives, such as object and state attributes, as new combinations are introduced to the model.
HOW THIS AFFECTS YOU
●
researcherThis offers a parameter-efficient way to manage class-incremental learning without losing prior knowledge of primitives.