https://jaehun.me/en/tags/world-models/Recent content in World Models on Jaehun's BlogHugoen-USThu, 15 Oct 2026 14:17:42 +0900https://jaehun.me/en/posts/paper-2610-05739/Thu, 15 Oct 2026 00:00:00 +0900https://jaehun.me/en/posts/paper-2610-05739/<p><a href="https://arxiv.org/abs/2610.05739"target="_blank" class="inline-flex items-center gap-1" >Paper<svg class="h-3 w-3 flex-shrink-0" id="external-link" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15 3h6v6m-11 5L21 3m-3 10v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/></svg> </a></p> <p class="paper-source">arXiv:2610.05739 · Monash University · The University of Adelaide · Zhejiang University</p> <h2 id="one-line-summary-tldr">One-Line Summary (TL;DR)<a href="#one-line-summary-tldr" class="heading-anchor" aria-label="Link to this section"></a></h2><p>Recurrent linear attention (Gated DeltaNet) in video world models compresses memory into a fixed-size state, which is efficient, but when the camera leaves and returns, distant scenes are progressively forgotten. <strong>HLA-WM</strong> caches chunk-level affine summaries in a training-free way and uses <strong>camera geometry (frustum overlap)</strong> to retrieve and recompose the relevant scenes, achieving <strong>PSNR +0.74 dB</strong>, <strong>rotation error −28.5%</strong>, and a <strong>12× reduction versus Full KV</strong> in history memory over a 60-second rollout (about 6× the memory of GDN) (source: §Abstract, Fig. 1).</p>