HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation

arXiv 2026 56 method, application

TLDR

HorizonDrive introduces a self-corrective autoregressive world model for long-horizon driving simulation via scheduled rollout recovery and teacher rollout DMD.

Reasoning

The paper presents a novel training-and-distillation framework that addresses drift in autoregressive rollout by making the teacher rollout-capable, enabling unbounded-horizon supervision. Strengths include a clear problem formulation and a practical solution for long-horizon simulation, but the abstract lacks explicit real-world validation or empirical benchmarks, limiting evidence of practical impact.

Read-first score

Read-first score 56, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Recency 6%

100

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

Citation impact 18%

77.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.77700905

Topical relevance 29%

75.7

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

Reproducibility 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 463.

Keyword Scores

world model

generative world model

world dynamics prediction

world simulator

interactive world model

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Scheduled rollout recovery (SRR) trains the base model to reconstruct ground-truth future clips from prediction-corrupted histories, yielding a teacher that remains stable across long AR rollouts.
Teacher rollout DMD (TRD) enables a short-window student to align with the rollout-capable teacher for efficient real-time deployment.

Methodology

HorizonDrive is an anti-drifting training-and-distillation framework for autoregressive driving simulation. First, scheduled rollout recovery (SRR) trains a base model to reconstruct ground-truth future clips from prediction-corrupted histories, producing a teacher stable across long AR rollouts. Then, the rollout-capable teacher is extended via AR rollout to provide long-horizon distribution-matching supervision under bounded memory, while a short-window student aligns to it using teacher rollout DMD (TRD) for efficient real-time deployment.

Key Results

On nuScenes, HorizonDrive reduces FID by 52% and FVD by 37%, and lowers ARE and DTW by 21% and 9% relative to the strongest long-horizon streaming baselines, while remaining competitive with single-pass driving video generators.

Resources

Paper