Unveiling and Addressing Pseudo Forgetting in Large Language Models
This addresses a specific problem in continual learning for AI researchers by revealing a hidden mechanism in model performance degradation, though it is incremental as it builds on existing forgetting mitigation efforts.
The paper identifies 'pseudo forgetting' in large language models, where performance drops on old tasks are due to instruction failures rather than lost capabilities, and shows restoration via simple interventions like adding correct rationales or meaningless suffixes, with their RGD-R framework reducing this issue while preserving plasticity.
Although substantial efforts have been made to mitigate catastrophic forgetting in continual learning, the intrinsic mechanisms are not well understood. In this work, we demonstrate the existence of "pseudo forgetting": the performance degradation on previous tasks is not attributed to a loss of capabilities, but rather to the failure of the instructions to activate the appropriate model abilities. We show that the model's performance on previous tasks can be restored through two simple interventions: (1) providing partial external correct rationale, and (2) appending semantically meaningless suffixes to the original instructions, to guide the generation of correct rationales. Through empirical analysis of the internal mechanisms governing rationale generation, we reveal that models exhibiting pseudo forgetting show reduced instruction dependence during rationale generation, leading to suboptimal activation of their inherent capabilities. Based on this insight, we propose Rationale-Guidance Difficulty based Replay (RGD-R) framework that dynamically allocates replay data based on the model's ability to correctly leverage the intrinsic capabilities. Experimental results demonstrate that RGD-R effectively mitigates pseudo forgetting while maintaining model plasticity.