Autoregressive world models excel at sample‑efficient decision making, yet their open‑loop rollouts tend to drift, causing error accumulation even within modest planning horizons. Hierarchical temporal abstraction can shorten the effective prediction horizon to curb drift, but existing hierarchies often add structural and conceptual complexity, making them hard to control and limiting gains over flat baselines.\ \ MHWA (Multi-timescale Hierarchical World‑Action model) introduces a strided high‑level context channel that reduces the update frequency of global context, thereby limiting the propagation of compounding drift. The core innovation is context‑aware routing: a gating network dynamically reconfigures a low‑level mixture of experts using both global context and current observations, enabling flexible high‑low coordination and alleviating interference typical of static hierarchies.\ \ Empirically, with RLiable’s seed‑by‑game aggregation, MHWA (77 M parameters) attains an IQM human‑normalized score of 0.884 across three training seeds on a 14‑game Atari subset using 10% subsampled offline data. This outperforms a matched JOWA‑70 M baseline (0.529) and matches the performance of JOWA‑150 M while using roughly half the parameters and 70.88% fewer training FLOPs per sequence.\ \ Review