E²-LoRA: Energy-Ordered Low-Rank Adaptation for Continual Learning

A new method called E²-LoRA (Energy-Concentrated and Energy-Ordered Low-Rank Adaptation) is proposed to address task interference in Continual Learning (CL). The approach is based on the observation that output feature drift from parameter updates is inherently low-rank, and preserving parameters along principal directions minimizes reconstruction error. E²-LoRA orders and concentrates knowledge into leading ranks, freeing capacity for future tasks. A dynamic rank allocation strategy balances stability and plasticity by optimizing energy retention and model plasticity. The method is validated across multiple benchmarks.