⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Memorizon: Training World Models Beyond Their Context Window

arXiv 2026 34.1 method

TLDR

Memorizon trains streaming world models on arbitrarily long spans by retrieving top-K latents via camera co-visibility, enabling consistent repeated-visit rendering without quadratic attention cost.

Reasoning

The paper presents a novel training recipe that decouples long-span supervision from attention cost, with empirical evidence of improved revisit consistency and bounded step-time growth. Weaknesses include a trade-off with image quality and reliance on camera co-visibility, with no demonstrated interactive or RL settings.

Read-first score

Read-first score 34.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Recency 5%
100

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

Methodology quality 15%
60

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

Topical relevance 25%
54.3

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

Reproducibility 15%
38

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

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%
0

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

Citation velocity 10%
0

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 280.

Keyword Scores

world model
10
video world model
8
generative world model
7
world simulator
6
world dynamics prediction
5
interactive world model
1
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • Decouples long-span supervision from attention: training samples can cover arbitrary spans but are scored only on the last k chunks, so revisits can share a forward pass without including every intervening frame.
  • Introduces a retrieval-based shared memory bank bounded by kK: each scored chunk retrieves its own top-K latents by camera co-visibility, and the union of these requests forms the bank, keeping sequence length bounded regardless of span.
  • Scales training to long spans with minimal overhead: adding the bank raises step cost once, and extending span from 100 to 400 seconds adds only 12% step time.

Methodology

Memorizon trains streaming world models on samples that cover spans of any length but are scored only on the last k chunks. Each scored chunk retrieves top-K latents by camera co-visibility, and the union of these requests forms a shared bank bounded by kK. The approach is compared against a sliding-window baseline and evaluated on revisit consistency and image quality under varying span lengths and bank sources.

Key Results

Retrieval raises revisit consistency on every split over a sliding-window baseline. Spans long enough to reach the first visit of each return add a further 24% to 30% consistency but at some cost in image quality; beyond that span, more length no longer helps. Filling the bank from another episode lowers revisit correlation by 83%.

Limitations

  • The gains in revisit consistency come at some cost in image quality.
  • Extending the span beyond the first visit of each return does not further improve revisit consistency.

Resources

Paper

Tags