Catastrophic Forgetting
Catastrophic forgetting is a sharp loss of previously learned capability when a model trains on new data or tasks. The skill is recognizing interference, measuring retained competence and choosing adaptation methods that balance learning the new requirement with preserving behaviors that remain important to the application.
What it is
Neural network parameters are shared across tasks, so updates that reduce a new task's loss can disrupt representations or decision boundaries needed for earlier tasks. This differs from a change in the evaluation population: forgetting compares capability on the same earlier requirement before and after adaptation. Mitigation families include replaying old examples, constraining changes to important parameters, separating task-specific components and balancing data mixtures. Elastic Weight Consolidation illustrates regularization based on parameter importance. No single mechanism defines all forgetting, and the acceptable balance depends on whether older tasks, domains, languages or safety behaviors must still be supported.
What the work involves
Establish a baseline on old and new tasks before sequential training. Keep representative retention sets separate from replay data, and track performance after each adaptation stage. Compare targeted replay, smaller updates, regularization and isolated adapters when they fit the model. Group evaluation examples by original source to prevent a remembered template from overstating retention. Report the new capability gain alongside losses elsewhere. A useful deliverable is an adaptation strategy and checkpoint selection rule that makes the retention tradeoff explicit, with rollback criteria for critical behaviors.
Illustrative example
An illustrative multilingual classifier is adapted to specialist English support tickets. The developer measures its original French and Polish tasks after each training stage. New-ticket accuracy improves while earlier language performance declines. They compare a mixed-language replay set with an isolated specialist adapter and evaluate both on untouched tickets from each language. The selected approach supports the new domain while meeting the project's minimum retained capability, rather than optimizing English performance alone.
Limits and common mistakes
Replay requires access to suitable older data and may preserve its biases. Importance estimates and parameter constraints are approximations, while separate adapters add routing and maintenance decisions. Retention tests can overlook rare skills or reward superficial similarity to old answers. Domain shift and forgetting can coexist, so diagnose them separately. Forgetting is a measured failure mode, not a guarantee that any particular fine-tuning recipe will fail; verify before and after behavior under the same evaluation protocol.
Prerequisites
- hardLLM Fine-Tuning
Catastrophic forgetting IS a fine-tuning problem — it only occurs during fine-tuning or CPT
Related skills
- → is part of: Model Fine-Tuning
Sources and further reading
- Overcoming catastrophic forgetting in neural networks
Sequential task interference and parameter-importance regularization as one mitigation.
- Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks
Domain adaptation context for distinguishing new-domain gains from retention requirements.
Last updated: 2026-10-10