HLA-WM: training-free hybrid attention gives video world models 12x memory savings

Monash · hf · 2026-10-06

Researchers propose HLA-WM, a training-free hybrid linear-attention framework for long-horizon video world models:

Motivation: long rollouts need persistent memory, but recurrent linear attention (Gated DeltaNet) suffers severe long-range forgetting as later state updates attenuate distant scenes.
Method: exploit GDN's affine structure to cache compact chunk-wise transition summaries, retrieve scene-relevant historical chunks via camera geometry, and recompose them into query-specific recurrent states — combining coarse geometry-guided retrieval with fine-grained linear-state computation.
On the 60-second SANA-WM-Bench, improves all six metrics with no extra training (+0.74 dB PSNR, 28.5% lower rotation error), generalizing to MBench-A.
At 60-second context, cuts historical-state memory 12x vs full KV caching with at most 1.6% throughput loss.

More from Embodied