Code, results and checkpoints for the paper "Silent Pipeline Defects in World-Model RL: A Dreamer 4 Case Study": a PyTorch reimplementation of Dreamer 4, two silent defects that passed 12 checks, and three cheap checks that catch them
reinforcement-learning pytorch transformer reproducibility model-based-rl dreamer world-models offline-rl world-model flow-matching dreamer-v4 imagination-training pmpo
-
Updated
Oct 22, 2026 - Python