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

LongDWM: Cross-Granularity Distillation for Building a Long-Term Driving World Model

arXiv 25.6 2025 69.1 method

TLDR

A long-term driving world model using hierarchical decoupling and cross-granularity distillation to reduce error accumulation and improve video coherence.

Reasoning

The paper addresses a clear limitation (error accumulation in long-term video prediction) with a novel hierarchical distillation approach, achieving significant improvements on NuScenes. Strengths include strong empirical results and efficiency gains; weaknesses are limited to front-view only and lack of interactive or RL context.

Read-first score

Read-first score 69.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Reproducibility 25%
73

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

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%
60

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 78.

Keyword Scores

world model
10
video world model
10
generative world model
9
world dynamics prediction
9
world simulator
7
interactive world model
2
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • Hierarchical decoupling of world model learning into large motion learning and bidirectional continuous motion learning
  • Distillation method where fine-grained video flows serve as self-supervised signals for coarse-grained flows
  • Coordination of coarse-grained and fine-grained modules to generate long-term temporally coherent videos

Methodology

The model hierarchically decouples world model learning into large motion learning and bidirectional continuous motion learning. It proposes a distillation method where fine-grained video flows act as self-supervised signals for coarse-grained flows, and the coarse-grained and fine-grained modules are coordinated to generate long-term and temporally coherent videos.

Key Results

On the NuScenes benchmark, compared with the state-of-the-art front-view model, the proposed model improves FVD by 27% and reduces inference time by 85% for generating 110+ frames.

Limitations

  • Current models suffer serious error accumulations when predicting the long-term future
  • Training on short video clips (high fps and short duration) leads to a training-inference gap
  • Multiple roll-out generations struggle to produce consistent and reasonable long videos

Resources

PaperCode

Tags