Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay
2026-05-25 • Machine Learning
Machine Learning
AI summaryⓘ
The authors studied how computer models forget old tasks when learning new ones, a problem called forgetting. They found that instead of storing old data, models can create their own examples to remember past tasks, which helps prevent forgetting. However, if the model is already very full of knowledge, it still struggles to learn new things without losing old ones. Using slow learning can avoid forgetting but takes longer, while creating replay data lets the model learn quickly without forgetting.
forgettingreplaylanguage modelstraining distributioncapacitylearning ratefinetuningexemplarscatastrophic forgetting
Authors
Martin Marek, Dongkyu Cho, Shikai Qiu, Rumi Chunara, Pavel Izmailov, Andrew Gordon Wilson
Abstract
Models trained on a new task typically degrade on prior tasks, a phenomenon known as forgetting. Traditionally, mitigating forgetting has required replaying stored exemplars from prior tasks, which is often impractical. By contrast, language models can sample from their own training distribution, and we show that these self-generated samples serve as effective replay data, nearly eliminating forgetting. We find that forgetting nonetheless persists when the model has little remaining capacity: models pretrained close to saturation cannot absorb new information without overwriting prior knowledge. When capacity is not the limiting factor, low learning rates reduce forgetting but require substantially more training steps. Replay breaks this tradeoff, enabling fast, high-learning-rate finetuning without forgetting.