Orbis 2: A Hierarchical World Model for Driving

arXiv 2026 47.2 method, application

TLDR

A hierarchical driving world model with two-level prediction and diffusion forcing pretraining achieves state-of-the-art results on benchmarks.

Reasoning

The paper introduces a novel hierarchical world model for driving that factorizes prediction across two abstraction levels, and a two-stage training paradigm combining diffusion and teacher forcing. Strengths include clear methodology, strong empirical results on standard benchmarks, and improved internal representations. Weaknesses are domain specificity to driving and lack of explicit discussion of limitations or generalizability.

Read-first score

Read-first score 47.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 6%

100

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

Methodology quality 18%

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

Topical relevance 29%

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

Reproducibility 18%

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

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

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 499.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Hierarchical world model with two levels: high-level predictor for coarse scene structure over long horizons, and low-level generator for detailed predictions conditioned on high-level output.
Diffusion forcing objective for pretraining, yielding richer internal representations compared to teacher forcing.
Two-stage training paradigm: pretrain with diffusion forcing, fine-tune with teacher forcing to combine representational benefits and rollout stability.
State-of-the-art performance on driving world model benchmarks including long-horizon generation, steering responsiveness, and representation quality.

Methodology

A hierarchical world model factorizes future prediction into high-level (coarse, long-horizon) and low-level (detailed, conditioned) components. Training uses a two-stage approach: diffusion forcing pretraining followed by teacher forcing fine-tuning. The model is evaluated on standard driving world model benchmarks for generation fidelity, steering responsiveness, and internal representation quality.

Key Results

The method achieves state-of-the-art results on long-horizon generation fidelity, steering responsiveness in counterfactual scenarios, and internal representation quality across established driving world model benchmarks.

Resources

Paper