LOCI: Spatial Linear Memory for Streaming World Models

arXiv 2026 31.6 method

TLDR

LOCI is a hybrid spatial-memory architecture for streaming video world models, combining key-value caches and camera-conditioned recurrent linear attention to reproduce revisited scenes with lower memory.

Reasoning

The paper introduces a novel hybrid memory design that integrates projective camera geometry into recurrent linear attention, with strong empirical results on a public benchmark and real recorded trajectories. However, the abstract provides limited detail on baselines, ablations, and limitations, and the connection to broader world-model terminology is only partially explicit.

Read-first score

Read-first score 31.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 25%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=reproduce

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 238.

Keyword Scores

video world model

world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Hybrid spatial-memory architecture that keeps both key-value cache and recurrent linear-attention memory
Recurrent linear-attention memory with reads and writes conditioned on projective camera geometry
Alternating transformer blocks: cache-backed attention over past observations in half of blocks, chunk-restricted attention plus recurrent memory in the other half
Recurrent readouts flow into subsequent cache-backed blocks to supply queries with accumulated scene context
Bounded bank of retained observations enabling constant-memory streaming of long videos

Methodology

LOCI is a transformer-based video world model evaluated on the public MIND memory benchmark and held-out recorded trajectories. It uses a hybrid design where half of transformer blocks maintain a key-value cache of past observations, while the other half restrict attention to the current chunk and are complemented by a recurrent linear-attention memory conditioned on projective camera geometry. Comparisons are made against representative world models and a same-recipe full-softmax model using faithfulness on revisited content and memory usage.

Key Results

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, and with a bounded bank of retained observations it streams long videos at constant memory while remaining more faithful than full softmax under the same budget.

Resources

Paper