ReWorld: An Interactive World Model with Long-Horizon Memory

arXiv 2026 44.9 method

TLDR

ReWorld is an interactive world model with long-horizon memory, using mixed per-head attention, bounded KV cache, landmark retrieval, and distillation for real-time video generation.

Reasoning

The paper introduces a novel architecture separating short-horizon control from long-horizon memory, supported by a metric-scale-aligned multi-source data engine and distillation for real-time inference. Strengths include addressing a structural trade-off and evaluating against six recent models, but the abstract is truncated and lacks explicit quantitative results.

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=baseline,metric

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

interactive world model

generative world model

video world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Mixed per-head attention windows with a small set of global heads, random head routing, and random chunk dropping to decouple control and memory during training
Inference-time bounded KV cache backed by a pose-indexed landmark bank that retrieves the nearest landmarks to the current pose for long-horizon memory
Metric-scale-aligned data engine that places eight diverse sources (Unreal renders, game roaming, real-world footage) on a unified physical action scale, with palindrome trajectories for revisit evidence
Distribution-matching distillation confined to a LoRA adapter, compressing sampling to four steps and enabling one backbone to serve both a high-fidelity multi-step mode and a real-time interactive streaming mode

Methodology

ReWorld separates control and memory during training using mixed attention windows: most heads attend to a short recent window while a few global heads attend over the full history, with random head routing and chunk dropout to prevent specialization. At inference, it uses a fixed-budget KV cache supported by a pose-indexed landmark bank that retrieves relevant historical landmarks, bounding memory. A metric-scale-aligned data engine unifies eight data sources onto a common physical action scale and generates palindrome trajectories for revisit-based memory training. Finally, distribution-matching distillation into a LoRA adapter reduces sampling to four steps, enabling a single backbone to operate in both high-fidelity and real-time interactive modes.

Key Results

ReWorld achieves the best control fidelity (11.95° rotation error, best camera-motion consistency) and best generation quality against six recent interactive world models. On minute-long out-and-back rollouts (64 s, 384 latents) with a fixed 12-chunk cache, it regenerates the starting view, while sliding-window baselines evict the evidence and full-KV attention runs out of memory.

Resources

Paper