Catastrophic Forgetting — AI Glossary

Reference details

Related terms

• Continual Learning
• Adapter (PEFT)
• LoRA (Low-Rank Adaptation)

How it works

1. Fine-tuning computes gradients only on the new task, so weights encoding prior capability are updated with no signal to preserve them.
2. The optimiser has no notion of which weights matter for anything other than the current loss, so prior behaviour is overwritten rather than traded away deliberately.
3. Narrow data makes it worse: a small, homogeneous fine-tuning set pulls the model hard toward a distribution unlike its pretraining.
4. Parameter-efficient methods mitigate it structurally by freezing the base and training a small delta, so there is less that can be overwritten.
5. Replaying a fraction of general or prior-task data during the fine-tune is the most direct mitigation, because it restores the missing gradient signal.

Example: Why a small replay fraction recovers most of it

Fine-tune on 10,000 narrow examples and general capability degrades. Mix in 1,000 general examples — a 10% increase in training data — and the optimiser now receives gradient signal for both distributions, which typically recovers much of the loss. The mitigation is cheap in data and trivial in code; the reason it is skipped is that the regression is invisible unless the old tasks are still being evaluated, and most fine-tuning suites only cover the new one.

Commonly confused with

Treated as a synonym for continual learning, which is the field studying it rather than the phenomenon. This is the failure: new training destroys old capability. Continual learning is the discipline of avoiding it across a sequence of tasks. It is also not overfitting, though both come from narrow data — overfitting is poor generalisation on the new task itself, while this is degradation on tasks you stopped measuring. A model can fit its new task excellently and still have forgotten everything else, which is why the two need separate evaluation.

When to use it

Reach for it when:

• Planning a fine-tune where general capability must survive.
• Explaining why prior-task evaluation belongs in the acceptance criteria.
• Justifying parameter-efficient methods or data replay over full fine-tuning.

Reach for something else when:

• Describing poor generalisation on the new task, which is overfitting.
• Where the model is purpose-built for one narrow task and prior capability is irrelevant.
• As an argument against fine-tuning generally, since the mitigations are well understood.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Catastrophic Forgetting — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path