Generative Replay

Generative replay studies how AI systems can practice on synthetic examples of earlier tasks while adapting to new ones, reducing catastrophic forgetting when storing old data is impractical or restricted. Approaches include diffusion and other generative models, latent or statistical generators, and world models that imagine trajectories; replay is often combined with regularization or methods that select the most useful examples. The approach is used in settings such as continual vision and audio learning, language-model fine-tuning, and robot control, with evaluations ranging from standard benchmarks to real-world robots. Results can approach experience replay with stored data, but generated examples may drift from past experience or contain unrealistic observations and actions, making replay quality and validation central challenges.