How do LLM systems handle catastrophic forgetting during fine-tuning?

Catastrophic forgetting occurs when fine-tuning causes models to lose previously learned capabilities, and mitigation techniques preserve general knowledge during adaptation.

Fine-tuning specializes models for specific domains, but aggressive updates may overwrite existing representations.

This phenomenon is called catastrophic forgetting.

For example:

Mitigation strategies include:

Updating only small adapter layers.

Injecting trainable low-rank matrices.

Combining old and new datasets.

4.…