OmniDrive: An LLM-Choreographed Multi-Agent World Model with Unified Latent Co-Compression for Multi-View Driving Video Generation

arXiv 2026 64.7 method, application

TLDR

An LLM-choreographed multi-agent world model for controllable multi-view driving video generation using unified latent co-compression.

Reasoning

The paper introduces a novel architecture (DRIVE-CHOREO) that uses multiple LLM agents to coordinate multi-view video generation, achieving state-of-the-art consistency and downstream utility. Strengths include addressing key tensions in driving world models and strong empirical results on nuScenes; weaknesses are limited to a single dataset and no explicit interactive or RL-based evaluation.

Read-first score

Read-first score 64.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.

Recency 6%

100

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

Citation impact 18%

95.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95274583

Methodology quality 18%

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

Topical relevance 29%

67.1

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=code,dataset

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 423.

Keyword Scores

generative world model

world model

video world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

LLM-choreographed multi-agent world model with three specialized agents (Director, Cartographer, Auditor) for controllable multi-view driving video generation
Unified latent co-compression via view-time permutation that enforces inter-camera geometry within the convolutional receptive field of a 3-D VAE
Shared symbolic interlingua (WorldScript) aligning language, geometry, and pixels at the latent-token level to resolve heterogeneous control injection and post-hoc cross-view fusion

Methodology

The method uses three Qwen2.5-VL agents: a Director that parses user intent into a structured WorldScript, a Cartographer that grounds it into spatially-anchored layout tokens, and an Auditor that feeds cross-view critiques as auxiliary supervision. These agents jointly author a position-aware token sequence, which is co-compressed with multi-view video via a view-time permutation in a 3-D VAE. The model is trained and evaluated on the nuScenes dataset, with baselines including state-of-the-art methods and metrics such as multi-view consistency, BEV mAP, FVD, and NDS.

Key Results

DRIVE-CHOREO achieves state-of-the-art multi-view consistency and BEV mAP of 21.6 with competitive FVD of 45.7 on nuScenes. A detector trained purely on synthetic data from the model gains +2.4 NDS on the real validation split, demonstrating downstream utility.

Resources

Paper