Researchers have developed GR4CIL, a new framework designed to improve Class Incremental Learning (CIL) for CLIP-based models. This approach addresses challenges in adapting shared parameters and organizing task-specific knowledge, which often lead to knowledge drift and poorly calibrated responses. GR4CIL enhances task discrimination and knowledge routing by preserving specific visual knowledge and stabilizing the shared textual semantic space, while an orthogonal compensation mechanism mitigates modality gaps and improves score margins. Experiments demonstrate that GR4CIL surpasses existing methods on multiple benchmarks. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
RANK_REASON This is a research paper detailing a new framework for class incremental learning.