Video World Models · 2026-10-01
LOCI: Spatial Linear Memory for Streaming World Models
Ji Xia, Tingting Liao, Xuezhi Liang, Hao Li, Guangyi Liu
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
LOCI is a streaming video world model for revisits. It uses a hybrid spatial memory: half the transformer blocks keep a key–value cache of past observations, while the other half uses a projectively conditioned recurrent linear-attention memory with camera-geometry-aware reads/writes. Recurrent readouts shape subsequent historical queries. The authors report better revisit fidelity than representative world models and a same-recipe full-softmax, and lower peak memory with full history (~30%).
Why it matters
Scene revisits require retrieving the right past observations for the current viewpoint without letting memory grow with video length. The authors’ hybrid keeps both observation-level detail (KV cache) and compact long-term integration (recurrent state), using camera geometry to address memory. They evaluate on the public MIND memory benchmark and held-out recorded trajectories, including bounded streaming with constant memory.
Method
- Hybrid memory across transformer blocks: main attention keeps a key–value cache of past observations in half the blocks; the other half uses recurrent Kimi Delta Attention (linear-attention) read/write conditioned on projective camera geometry.
- Chunk-synchronous memory dynamics: hybrid blocks read the preceding fixed-size state; tokens write/update the state via a delta rule at token resolution; retention is applied once per chunk.
- Dense vs. bounded historical access: dense softmax attends over the available prefix; bounded sparse keeps a recent window plus a fixed-capacity bank of older views while the recurrent state carries full-history information so generation can stream at constant memory.
Limitation
The paper text does not state a specific limitation of the method.
Abstract (from arXiv)
When a camera revisits a previously observed region, a video world model should reproduce what was there before. This requires both remembering past observations and retrieving the right one for the current viewpoint. Key-value caches preserve visual detail but grow with video length; recurrent memory is compact but compresses history into a fixed-size state, so individual past observations are no longer directly accessible. We introduce LOCI, a hybrid spatial-memory architecture that keeps both representations. In half of the transformer blocks, main attention keeps a key-value cache of past observations; in the other half, it is restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, so viewpoint enters both memory addressing and stored content. Recurrent readouts flow into subsequent cache-backed blocks and supply their queries with accumulated scene context. On the public MIND memory benchmark and on held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model; with full history, it lowers peak memory at equal length by about 30% relative to full softmax. With a bounded bank of retained observations, it streams long videos at constant memory and remains more faithful than full softmax under the same budget.
Related papers
- Honeycomb: Constant-Size Scene Memory Representation for Video World Models
- HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models
- Learning to Retrieve: Internalizing Memory Retrieval for Video World Models
- World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
- Persistent Planar Memory for Video World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-09-30 · Indexed by Pub-AI: 2026-10-01 UTC
LOCI: Spatial Linear Memory for Streaming World Models
Preprint posted: 2026-09-30 · Original record ↗
- arxiv
- 2609.40222
Explicit source versions (1)
- v1 · 2026-09-30
- Preprint available
Source date: 2026-09-30 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
When a camera revisits a previously observed region, a video world model should reproduce what was there before. This requires both remembering past observations and retrieving the right one for the current viewpoint. Key-value caches preserve visual detail but grow with video length; recurrent memory is compact but compresses history into a fixed-size state, so individual past observations are no longer directly accessible. We introduce LOCI, a hybrid spatial-memory architecture that keeps both representations. In half of the transformer blocks, main attention keeps a key-value cache of past observations; in the other half, it is restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, so viewpoint enters both memory addressing and stored content. Recurrent readouts flow into subsequent cache-backed blocks and supply their queries with accumulated scene context. On the public MIND memory benchmark and on held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model; with full history, it lowers peak memory at equal length by about 30% relative to full softmax. With a bounded bank of retained observations, it streams long videos at constant memory and remains more faithful than full softmax under the same budget.