Computer Science

MASTERING MEMORY TASKS WITH WORLD MODELS

M. R. Samsami, A. Zholus, et al.

Model-based RL agents struggle with long-term dependencies—Recall to Imagine (R2I) fixes this by integrating a new family of state space models into world models to boost long-term memory and long-horizon credit assignment. R2I sets new state-of-the-art on memory and credit-assignment benchmarks like BSuite and POPGym, achieves superhuman results on Memory Maze, matches performance on Atari and DMC, and converges faster than DreamerV3. This research was conducted by Mohammad Reza Samsami, Artem Zholus, Janarthanan Rajendran, and Sarath Chandar.