Replay on Demand lets continued-pretraining models decide when forgotten knowledge deserves replay

Replay on Demand replaces a fixed replay mixture with a shared competition between adaptation and retention examples, using model-dependent scores to allocate training compute as forgetting emerges.