HLA-WM Hybrid Linear Attention for Long-Horizon Video World Models
October 4, 2026
HLA-WM addresses long-range forgetting in Gated DeltaNet models by combining geometry-guided retrieval with recurrent linear-state computation. This training-free framework uses chunk-wise transition summaries to preserve scene consistency during extended video rollouts without the memory overhead of full KV caches.
HOW THIS AFFECTS YOU
●
builderYou can achieve better scene consistency in video generation with lower memory costs than standard softmax attention.
●
researcherYou can leverage the affine structure of GDN to improve long-range dependency in recurrent architectures.