From Weights to Features: Sparse-Autoencoder-Guided Activation Regularization for Continual Learning in Large Language Models
Evan Ning, Wei Xue, Dong Lou, Yike Guo
原文摘要
Weight-protecting regularizers such as elastic weight consolidation (EWC) are a standard defense against catastrophic forgetting; however, in the sequential fine-tuning of large language models, they exhibit a stability-plasticity tradeoff that penalty tuning cannot resolve. We attribute this limitation in part to polysemanticity: because each weight participates in many unrelated concepts, a per-weight importance score has difficulty isolating the knowledge that requires protection. We instead regularize in activation space, using a pretrained sparse autoencoder (SAE) as a fixed dictionary of approximately monosemantic features. We formulate continual learning as constrained optimization over feature drift, with a stability constraint on the features to be preserved and a plasticity constraint on the features used by the current task, and we show that the resulting squared-hinge loss is a regularized Lagrangian relaxation of this problem, with EWC as a special case. A replay variant measures drift on 200 stored sequences per previous task, whereas a data-free variant stores only a 197 KB feature mask per task, versus 6.5 GB for EWC. On TRACE (Gemma-2 9B, Llama-3.1 8B) and MedCL (Gemma-2 9B), the replay variant matches or exceeds SAE-based feature distillation at approximately one third of its per-step cost, and the data-free variant attains the highest mean overall performance among methods that store no previous-task data and add no task-specific parameters. Task-relevant units are linearly separable in the SAE basis but not in the neuron basis (ROC-AUC of 0.87-0.88 versus 0.50), and weight-space protection constrains the features of other tasks nearly as strongly as those of the protected task.
cs.LG · cs.CL