activation-based

AWARe is a fine-tuning method that mitigates catastrophic forgetting in multimodal large language models by selectively freezing important parameters based on activation patterns, preserving upstream capabilities while adapting to downstream tasks.