Addressable Memory for Video World Models

arXiv 2026 44.9 method, system

TLDR

Proposes WorldTrace, a training-free memory framework for video world models to maintain addressable compressed memory for long-horizon visual persistence, improving temporal consistency and episodic recall on a new benchmark LoopBench.

Reasoning

The paper identifies a concrete problem in video world models (memory addressing beyond training horizon) and offers a novel solution with two compression methods, plus a new benchmark. Strengths include clear problem definition and practical improvements; weaknesses include narrow focus on memory and evaluation only on their own benchmark.

Read-first score

Read-first score 44.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 6%

100

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

Topical relevance 29%

78.6

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 18%

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

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

Citation velocity 12%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 477.

Keyword Scores

world model

video world model

interactive world model

generative world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

WorldTrace: a training-free memory framework for long-horizon visual persistence that assigns in-distribution virtual positions to compressed memory slots, keeping them addressable despite rollout lengths beyond the training horizon.
WorldTrace-Field: a compression approach that compresses history for temporal coherence.
WorldTrace-Landmark: a compression approach that stores verbatim scene traces at detected transitions for episodic recall.
LoopBench: a new benchmark for evaluating whether a compressed cache can reconstruct a previously visited scene after a long detour.

Methodology

The paper proposes WorldTrace, a training-free framework that mitigates the failure of KV cache retrieval in video world models when rollouts exceed training horizon due to RoPE offsets. It compresses the memory and assigns each summary slot a distinct virtual position within the training distribution, then studies two compression variants: WorldTrace-Field for temporal coherence and WorldTrace-Landmark for episodic recall. Evaluation is performed on the newly introduced LoopBench benchmark.

Key Results

WorldTrace-Field improves temporal consistency by +15.5% and WorldTrace-Landmark improves episodic recall by +19.5% on LoopBench, extending visually persistent generation without retraining.

Resources

Paper