MosaicMem: Hybrid Spatial Memory for Controllable Video World Models

arXiv 26.3 2026 61.4 method

TLDR

MosaicMem hybrid spatial memory improves pose adherence and dynamic modeling in controllable video world models via 3D patch lifting and patch-and-compose interface.

Reasoning

The paper introduces a novel hybrid memory approach that addresses key limitations of explicit and implicit spatial memories for video world models, with clear methodology and experimental validation. However, the abstract lacks details on datasets and quantitative results, and the connection to model-based reinforcement learning is absent.

Read-first score

Read-first score 61.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 8%

100

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

Topical relevance 42%

68.6

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 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 326.

Keyword Scores

video world model

world model

world simulator

interactive world model

generative world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Hybrid spatial memory combining explicit 3D patches for reliable localization and targeted retrieval with implicit memory for dynamic modeling
Patch-and-compose interface that composes spatially aligned patches in the queried view
PRoPE camera conditioning for improved pose adherence
Two new memory alignment methods

Methodology

MosaicMem lifts patches into 3D for reliable localization and targeted retrieval, while exploiting the model's native conditioning to preserve prompt-following generation. It composes spatially aligned patches via a patch-and-compose interface, preserving persistent elements and allowing the model to inpaint evolving content. The method uses PRoPE camera conditioning and two new memory alignment methods, and is evaluated against implicit and explicit baselines.

Key Results

Experiments show improved pose adherence compared to implicit memory and stronger dynamic modeling than explicit baselines. MosaicMem enables minute-level navigation, memory-based scene editing, and autoregressive rollout.

Resources

Paper Project