arxivcs.LGcs.AI2026-07-22
The World Model Remembers, the Actor Forgets: Dream Rehearsal for Continual Model-Based RL
Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience. We ask a question the continual-RL literature has assumed an answer to but never meas…