Real Time Detection and Quantitative Analysis of Spurious Forgetting in Continual Learning

This paper addresses a critical challenge in continual learning for large language models: spurious forgetting. It moves beyond qualitative descriptions by introducing a quantitative framework to characterize alignment depth, identifying shallow alignment as a key vulnerability. The proposed framework offers real-time detection methods, specialized analysis tools, and adaptive mitigation strategies. The experimental results, demonstrating high identification accuracy and improved robustness, suggest a significant advancement in addressing spurious forgetting and promoting more robust continual learning in LLMs. The work's focus on practical tools and metrics makes it particularly valuable for researchers and practitioners in the field.

• •Introduces a quantitative framework for analyzing alignment depth in continual learning.
• •Provides real-time detection methods for identifying shallow alignment during training.
• •Demonstrates improved robustness against spurious forgetting through adaptive mitigation strategies.