Can World Models Benefit VLMs for World Dynamics?

arXiv 2025 59.8 method

TLDR

Investigates using video diffusion models as generative encoders in VLMs, finding improved spatial and multi-frame reasoning, achieving SOTA.

Reasoning

Strengths include systematic investigation of world model priors for VLMs and a novel method (DyVA) with empirical evidence. Weaknesses: abstract lacks details on limitations, generalizability, and computational cost.

Read-first score

Read-first score 59.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

61.4

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 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 197.

Keyword Scores

world model

generative world model

video world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Re-purposing a video diffusion model as a generative encoder (WorldLMs) for VLMs, using a single denoising step to produce visual embeddings
Dynamic Vision Aligner (DyVA) enabling enhanced spatial reasoning and multi-frame reasoning from single images
Curated suite of visual reasoning tasks for systematic evaluation of world model priors in VLMs

Methodology

The paper repurposes a video diffusion model as a generative encoder, performing a single denoising step to extract latent visual embeddings, which are then integrated into a Vision-Language Model (termed WorldLM). The best variant, DyVA, is trained and evaluated on a curated suite of visual reasoning tasks against open-source and proprietary baselines.

Key Results

DyVA surpasses both open-source and proprietary baselines on visual reasoning, achieving state-of-the-art or comparable performance, with significant gains in spatial reasoning and the ability to perform multi-frame reasoning from single images due to motion-consistency internalization from video pre-training.

Resources

Paper