Video World Models · 2026-10-06

HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models

Zhuokun Chen, Feng Chen, Xi Lin, Xiyu Wu, Jiahao He, Jianfei Cai, Bohan Zhuang

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

HLA-WM is a training-free hybrid linear-attention method for long-horizon video world models. It addresses long-range forgetting in Gated DeltaNet (GDN) by caching chunk-wise affine (Aᵢ,Bᵢ) transition summaries, retrieving scene-relevant chunks using camera geometry overlap, and recomposing them into a query-specific recurrent state. It improves long-range revisit consistency and camera control on SANA-WM-Bench and generalizes to MBench-A, with reduced memory and small throughput cost.

Why it matters

Recurrent linear-attention models compress history into fixed-size states but the authors identify severe long-range forgetting in GDN-based world models, harming revisits after long camera excursions. HLA-WM provides selectively addressable long-range memory without retraining, aiming to improve long-range scene recall while keeping GDN-like efficiency. The reported benchmark gains include improvements in PSNR and camera rotation error, along with memory reduction relative to full KV caching.

Method

Identify long-range forgetting in GDN-based video world models due to attenuation of early information through intervening recurrent state transitions.
Propose HLA-WM: cache affine chunk-wise GDN summaries (Aᵢ,Bᵢ), use camera-pose/intrinsics plus a depth estimate to retrieve geometrically overlapping historical chunks, then recomposes selected summaries into a query-specific recurrent state.
Apply selective recomposition only to the main GDN memory (camera-control and local convolutional states use original recurrent updates); no additional training or fine-tuning needed.

Abstract (from arXiv)

Long-horizon video world models require persistent memory to preserve scene consistency over extended rollouts. Softmax attention retains the full generation history through a growing KV cache, whereas recurrent linear attention compresses history into fixed-size states with substantially lower memory cost. However, we identify severe long-range forgetting in Gated DeltaNet (GDN), where information from distant but relevant scenes is progressively attenuated by subsequent state updates. To address this limitation, we propose HLA-WM, a training-free hybrid linear-attention framework that combines coarse-grained geometry-guided retrieval with fine-grained recurrent linear-state computation. HLA-WM exploits the affine structure of GDN to cache compact chunk-wise transition summaries, retrieve scene-relevant historical chunks using camera geometry, and recompose them into query-specific recurrent states. On the $60$-second SANA-WM-Bench, HLA-WM improves all six aggregate revisit-consistency and camera-control metrics of the base autoregressive generator without additional training, including a $0.74$ dB PSNR gain and a $28.5\%$ reduction in rotation error. The improvements persist after downstream refinement and generalize to MBench-A, where HLA-WM consistently improves all three revisit-consistency metrics across all four subsets and all evaluated inference modes over $547$ samples. At a $60$-second context, HLA-WM reduces historical-state memory by $12\times$ relative to full KV caching while incurring at most a $1.6\%$ reduction in inference throughput. These results demonstrate that selectively addressable recurrent memory can improve long-range scene recall while preserving the efficiency advantages of GDN. Project page: https://caesarhhh.github.io/hla-wm/

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-05 · Indexed by Pub-AI: 2026-10-06 UTC

preprint