When fine-tuning CLIP for new classes, you need to separately stabilize how the model extracts features (what it looks for) and how it combines them (how it makes decisions), not just update the whole model at once.
AREA solves catastrophic forgetting in CLIP-based class-incremental learning by decomposing classification into two stages: extracting visual/textual attributes and aggregating them. It stabilizes attribute extraction using geometric analysis on embedding spaces and learns task-specific experts for aggregation, preventing new classes from overwriting knowledge of old ones.