Source note2 min read
Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
Source
hugging_face_papers
arxiv.org/abs/2608.13546
Source
- Source name: hugging_face_papers (discovered via Hugging Face Daily Papers; primary record is arXiv)
- Source kind: paper (preprint, not yet peer-reviewed)
- Authors / organization: Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao
- Venue: arXiv, cs.CV
- Published: 2026-08-13
- Collected: 2026-08-15
- URL: https://arxiv.org/abs/2608.13546
- PDF: https://arxiv.org/pdf/2608.13546
- DOI: none
- Registry group: frontier_ai_labs (hugging_face_papers, research_aggregator role)
- Evidence role: research_aggregator (surfaced by HF trending list; underlying document is an unreviewed arXiv preprint)
- Language: en
- Country / region: UNKNOWN
Plain-Language Brief
The paper presents an interactive video world model that keeps track of a scene's state outside the model itself (an externally stored, camera-indexed memory) so it does not need to hold everything in context, aiming for linear rather than quadratic memory growth during long generation sessions.
What the Document Says
The system maintains a "persistent world state" externally through a camera-indexed bank, retrieving only relevant viewpoint information to prevent context bloat during extended sessions. Instead of a fixed teacher model, the authors design one optimized for long-horizon supervision using sparse attention that combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, targeting linear rather than quadratic scaling in memory requirements. These capabilities are distilled into a three-step student model without classifier-free guidance. The paper reports state-of-the-art results on WBench benchmarks and competitive results on VBench, generating 1.5-second video chunks in about 2.11 seconds on an H200 GPU at 384x640 resolution.
Key Concepts Mentioned
- Persistent, externally stored world state (camera-indexed memory bank)
- Sparse attention (chunk-wise grouping + distant-frame retrieval + linear-attention global state)
- Teacher-student distillation for long-horizon video generation
- WBench, VBench benchmarks
Claims Made by the Document
- Linear rather than quadratic memory scaling for long-horizon interactive video generation.
- State-of-the-art results on WBench; competitive results on VBench.
- Chunk generation latency of approximately 2.11 seconds per 1.5-second video chunk on an H200 GPU at 384x640 resolution.
Evidence Used
Benchmark evaluation on WBench and VBench; latency measurement on H200 GPU hardware. Full experimental tables not captured in this extract.
Methods / Approach
Teacher model built for long-horizon supervision via sparse attention (chunk-wise grouping, distant-frame retrieval, linear-attention global state); distilled into a three-step student model that skips classifier-free guidance.
Limitations Stated by the Authors
Not captured in the fetched abstract extract.
Limitations Noted by the Agent
Agent interpretation — not from the source.
This is an unreviewed arXiv preprint surfaced via Hugging Face's trending papers list, not a peer-reviewed publication. Benchmark comparisons (WBench, VBench) and the H200 latency figure are self-reported by the authors; no independent replication is available. classification_confidence is medium because only the abstract-level extract was reviewed, not the full paper.
Open Questions
- Full architectural details of the camera-indexed memory bank and its retrieval mechanism.
- Comparison baselines used for the "state-of-the-art on WBench" claim.
- Whether code/weights are released.
Notes for Future Q&A
May include interpretation. Labeled where it does.
Relevant to questions about long-horizon/streaming video world models and memory-efficient generation architectures. Should be cross-checked against other long-horizon world-model papers in the vault (e.g., PlayWorld, DreamX-Phi 1.0 collected the same day) when answering comparative questions.
Raw Abstract / Excerpt
"The system maintains a 'persistent world state' externally through a camera-indexed bank, retrieving only relevant viewpoint information to prevent context bloat during extended sessions... sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state... achieves linear rather than quadratic scaling in memory requirements... generates 1.5-second video chunks in approximately 2.11 seconds on an H200 GPU at 384x640 resolution."
