Answers are grounded in this paper and cite their sources.
Prompt-based continual learning is a promising direction that effectively leverages the capabilities of pre-trained models. These methods typically learn task-specific prompts and predict task identity to select the prompt during inference. However, task interference caused by prompt misselection significantly constrains their performance. Although several studies have attempted to mitigate this issue, they treat all samples equally, without specifically addressing the samples that are primarily affected by prompt misselection. In this paper, we stratify samples into three distinct categories according to their different responses to prompt misselection, including susceptible samples, refractory samples, and resilient samples. This stratification reveals that misselection-induced interference primarily manifests in susceptible samples, as only their classification correctness is influenced by prompt misselection, while other samples remain unaffected. This realization drives us to design a novel prompt-based approach called Confidence-Stratified Classifier Calibration (CoSC), which harnesses the complementary roles of all categories to safeguard susceptible samples. Specifically, we apply targeted optimization based on the unique properties of each category to cooperatively calibrate classifiers of both the prompt instruction part and the prompt selector, thereby protecting susceptible samples from dual perspectives. Extensive experiments show that CoSC achieves state-of-the-art performance across various benchmarks in class-incremental learning.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: